KR20220095988A - Apparatus and method for detecting a voice attack against the voice assistant service - Google Patents

Abstract

Disclosed are a server and method for detecting a generation of a speech attack for a voice assistant service and a device providing the service. The server providing the voice assistant service for detecting the speech attack for the device comprises: a communication interface; a storage part; and a processor. Therefore, the present invention is capable of preventing a complete speech attack from being generated.

Description

APPARATUS AND METHOD FOR DETECTING A VOICE ATTACK AGAINST THE VOICE ASSISTANT SERVICE

Various embodiments relate to a voice assistant service and a method for detecting and preventing a voice attack on a device providing the same.

As multimedia technology and network technology develop, a user can receive various services using a device. In particular, as voice recognition technology develops, a user may input a voice (eg, utterance) into a device and receive a response message according to the voice input through a service providing agent.

However, when a user plays audio in a home network environment including a plurality of devices, a voice attack is hidden in the reproduced audio. A problem in which a plurality of devices are controlled may occur.

In addition, an artificial intelligence model may be used to detect a hidden voice attack, but a voice attack is generated in an environment including an artificial intelligence model that detects a voice attack in order to avoid the voice attack being detected by the artificial intelligence model. , it is difficult to detect a hidden voice attack when it is hidden in audio, etc. Accordingly, before the voice attack hidden in the audio is completed, a technique for detecting and defending the voice attacks attempted in the step for generating the completed voice attack is required.

Various embodiments provide an apparatus for detecting a voice attack on a voice assistant service, comparing an input audio signal with pre-stored audio signals to identify similar audio signals, text converted from the input audio signal, and identified similar audio To provide an apparatus capable of detecting events determined to be a voice attack against a voice assistant service by comparing texts converted from signals, and performing an operation to prevent a completed voice attack from being generated, and an operating method thereof. can

As a technical means for achieving the above technical problem, according to an embodiment, a method for a server providing a voice assistant service to detect a voice attack on a device, the method comprising: receiving an input audio signal from the device; comparing the received input audio signal with a plurality of audio signals received from the device and stored in the server; identifying similar audio signals of the input audio signal from among the plurality of pre-stored audio signals based on a comparison result of the received input audio signal and the plurality of pre-stored audio signals; converting the received input audio signal into first text by performing Automatic Speech Recognition (ASR); obtaining transformed second texts from the identified similar audio signals; comparing the first text converted from the input audio signal with the second text converted from the similar audio signals; and determining whether the input audio signal is an audio signal for attacking the voice assistant service based on a comparison result of the first text and the second text; including, the plurality of pre-stored audio signals are provided from the device to the server before the input audio signal is received.

In addition, according to an embodiment, a server providing a voice assistant service for detecting a voice attack on a device includes: a communication interface for performing data communication with the device; a storage unit for storing a program including one or more instructions; and a processor executing one or more instructions of the program stored in the storage, wherein the processor controls the communication interface to receive an input audio signal from the device, and transmit the received input audio signal to the device compared with a plurality of audio signals received from and stored in the storage unit, and based on a comparison result of the received input audio signal and the plurality of pre-stored audio signals, the input from among the plurality of pre-stored audio signals Identifying analogous audio signals of an audio signal, performing Automatic Speech Recognition (ASR), converting the received input audio signal into a first text, and obtaining converted second texts from the identified analogous audio signals, and , comparing the first text converted from the input audio signal with the second text converted from the similar audio signals, and based on a comparison result of the first text and the second text, the input audio signal is It may be determined whether the audio signal is an audio signal for attacking the voice assistant service, wherein the plurality of pre-stored audio signals are provided from the device to the server before the input audio signal is received.

The server for detecting a voice attack according to an embodiment may include, based on whether a plurality of similar audio signals are input for a predetermined period, and the degree of transformation of texts converted from the input similar audio signals is greater than or equal to a predetermined standard, the hostile user By detecting an audio signal for generating a voice attack of In addition, the server may prevent a completed voice attack from being generated by performing operations for preventing the voice attack based on the voice attack detection result.

1A is a diagram for describing an operation of generating a voice attack and a server detecting a voice attack generating stage, according to an exemplary embodiment.
FIG. 1B is a diagram for explaining a step of generating a voice attack detected by a server according to an exemplary embodiment, and is a flowchart illustrating in detail the step of generating a voice attack of FIG. 1A .
2 is a block diagram illustrating a configuration of a server according to an embodiment.
3 is a flow diagram illustrating methods in which a server operates to detect the generation of a voice attack on a device, according to an embodiment.
4A and 4B are diagrams for explaining a method in which a server compares an input audio signal with a plurality of pre-stored audio signals using an audio comparison module according to an exemplary embodiment.
FIG. 5 is a diagram for describing a method in which a server compares first text converted from an input audio signal with second text converted from a plurality of pre-stored audio signals using a text comparison module, according to an exemplary embodiment.
6 is a diagram for explaining a method for a server to determine whether an input audio signal is an audio signal for an attack on a voice assistant service using an attack determination module, according to an exemplary embodiment.
7 is a diagram for describing a method for a server to determine an operation for preventing generation of a voice attack by using an attack response module according to an embodiment.
8 is a diagram for describing another embodiment in which a server operates to detect generation of a voice attack on a device according to an embodiment.
9 is a diagram for describing a method in which a server receives an input audio signal in a database and generates a plurality of pre-stored audio signals according to an exemplary embodiment.
10 is a flowchart illustrating a method in which a server detects generation of a voice attack and performs a device control operation or a defense operation of generating a voice attack based on a detection result, according to an exemplary embodiment.
11 is a block diagram of a client device according to an embodiment.

Terms used in this specification will be briefly described, and the present invention will be described in detail.

The terms used in the present invention have been selected as currently widely used general terms as possible while considering the functions in the present invention, but these may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

The singular expression may include the plural expression unless the context clearly dictates otherwise. Terms used herein, including technical or scientific terms, may have the same meanings as commonly understood by one of ordinary skill in the art described herein. Also, terms including an ordinal number such as 'first' or 'second' used in this specification may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

In the entire specification, when a part "includes" a certain element, this means that other elements may be further included, rather than excluding other elements, unless otherwise stated. In addition, terms such as “unit” and “module” described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software, or a combination of hardware and software.

Hereinafter, with reference to the accompanying drawings, the embodiments of the present invention will be described in detail so that those of ordinary skill in the art can easily implement them. However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

In the disclosed embodiment, a voice command refers to an audio signal including a command for controlling at least one of a client device and a peripheral device. The voice command may include a call command for calling the client device and a command for controlling at least one of the client device and a peripheral device.

In the disclosed embodiment, a voice attack refers to an adversarial audio signal recognized as a voice command in a client device. That is, the voice attack is not an audio signal by a general user who normally uses the voice assistant service, but may be an audio signal recognized as a voice command by the client device. In addition, the voice attack may be hidden in audio, video files, and the like. When a voice attack hidden in an audio or video file is reproduced in the environment of a general user who wants to use the voice assistant service, the user's client device may recognize the hidden voice attack as the user's voice command and operate erroneously.

In the disclosed embodiment, the adversarial user refers to a user who inputs and modulates a voice attack using the hostile user's own client device. The hostile user may repeatedly input and tamper with voice attacks recognized as voice commands in his/her voice assistant service usage environment until the voice attacks are recognized as voice commands desired by the hostile user. In addition, when a voice attack is recognized by the client device as a voice command desired by the hostile user, the hostile user may use the completed attack audio as a voice attack, and may hide the aforementioned audio and video files. When a voice attack hidden in an audio or video file is reproduced by a client device of a general user, which is another user described above, the general user's client device may erroneously recognize the hidden voice attack as a voice command.

The voice attack generation phase refers to operations in which input and modulation of a voice attack are repeated, performed by a hostile user. Since the hostile user repeatedly modulates a voice attack until the voice attack is recognized as a voice command desired by the hostile user, the server according to the disclosed embodiment detects the voice attack generation step, in which input and modulation of the voice attack are repeated operations. can do.

In the disclosed embodiment, the client device and the server can recognize the hostile user's voice attack as a voice command, so that it can be recognized whether the input voice command is a voice attack of the hostile user or a voice command of a general user using the voice assistant service. there may not be

Accordingly, whether the signal input to the client device is recognized as a voice attack of a hostile user or a voice command of a normal user who normally uses the voice assistant service is determined by the server performing operations according to the disclosed embodiments. You don't know until you detect the stage of generating a voice attack by performing it. Accordingly, in embodiments to be described later, a signal input to the client device will be described as an input audio signal.

The server may determine whether the audio signal input to the client device is a signal for an attack on the voice assistant service by performing operations of detecting the step of generating a voice attack according to the disclosed embodiments.

1A is a diagram for explaining a step of generating a voice attack and an operation in which a server detects a step of generating a voice attack, according to an embodiment.

Referring to FIG. 1A , a server 2000 according to an exemplary embodiment may be a server 2000 providing a voice assistant service. The server 2000 can detect a Voice attack generation phase 100 , which consists of a series of steps that an Adversarial user 150 attempts to generate a complete attack audio. have.

The hostile user 150 modulates and outputs a plurality of attack audios little by little to generate a completed attack audio, and the server 2000 performs Automated Speech Recognition (ASR) on the plurality of attack audios. Check the texts. Specifically, when the completed attack audio is converted into text by performing ASR on the server 2000, the hostile user 150 outputs the attack audio to be transcribed into the text desired by the hostile user 150 While doing so, the steps of checking the converted text from the output attack audio, modulating the output attack audio and re-outputting it may be repeated.

For example, the hostile user 150 may generate the initial attack audio in order to generate the completed attack audio that is transcribed into the text “Hi Bixby, open the front door” in the server 2000 .

The hostile user 150 may transmit the initial attack audio to the server 2000 via the client device and check text converted from the initial attack audio transmitted from the server 2000 to the client device. The hostile user 150 may generate the first attack audio 101 by modulating the initial attack audio when the text converted from the initial attack audio is not “Hi Bixby, open the front door”. In this case, in the first attack audio 101 , the waveform of the initial attack audio and the audio signal is similar, but the text converted from the first attack audio 101 and the text converted from the initial attack audio may be different. In the same way, when the text converted from the first attack audio 101 is not the text desired by the hostile user 150, the hostile user 150 modulates the first attack audio 101 to obtain the second attack audio 102 ) can be created. The hostile user 150 may repeat the above-described series of steps to generate the completed attack audio transcribed into the text “Hi Bixby, open the front door”.

The server 2000 according to an embodiment may detect the voice attack generation step 2000 before the attack audio completed by the hostile user 150 is generated. Specifically, a plurality of similar audio signals are input, and based on the degree to which texts converted from similar audio signals differ from each other, it can be determined as the voice attack generation step 100 attempted by the hostile user 150 . have.

FIG. 1B is a diagram for explaining a step of generating a voice attack detected by a server according to an exemplary embodiment, and is a flowchart for describing the step of generating a voice attack of FIG. 1A in detail.

In step S110, the hostile user 150 may output the initial attack audio through a speaker or the like. The initial attack audio output from the speaker may include a wake up command for calling the client device 1000 and a control command for controlling at least one of the client device 1000 and peripheral devices. When the initial attack audio is input to the client device 1000 , the client device 1000 may recognize the initial attack audio as a voice command and be activated.

In operation S120 , the client device 1000 may recognize the attack audio as a voice command, and a function for providing a voice assistant service for the recognized voice command may be activated. From the point of view of the client device 1000 , it is not possible to know whether a user inputting a voice command is the hostile user 150 or a general user who wants to use the voice assistant service. In this case, since the client device 1000 recognizes the attack audio of the hostile user 150 as a voice command, it transmits an audio signal representing the recognized voice command to the server, and sends a response to the voice command to the server 2000 . may be requested Also, an audio signal representing a voice command from the client device 1000 to the server 2000 may be stored in the audio signal database of the server 2000 to form a pre-stored audio signal database.

In step S130, the server 2000 according to an embodiment performs ASR on the received audio signal to convert it into text, and uses a Natural Language Understanding Model (NLU) to convert text from the received audio signal. can be analyzed. The server 2000 generates control information for controlling the client device 1000 or peripheral devices based on the analysis result, and includes text converted from the generated control information and the received audio signal. You can create a response. The server 2000 may provide a response to the voice command to the client device 1000 .

In operation S140 , the client device 1000 may receive a response to the voice command from the server 2000 and may perform an operation corresponding to the received response. However, since the client device 1000 recognizes the initial attack audio as a voice command and the server 2000 performs ASR on the audio signal corresponding to the initial attack audio, the converted text does not include clear semantic elements, so control information etc. may not have been generated, or control information unintended by the hostile user 150 may have been generated. Accordingly, the client device 1000 may not operate or may perform an operation not intended by the hostile user 150 .

In step S150 , the hostile user 150 may obtain the text converted from the attack audio included in the received response from the client device 1000 . In addition, the hostile user 150 may check the converted text to confirm whether the text desired by the hostile user 150 has been obtained (eg, “Hi Bixby, open the front door”).

When the text desired by the hostile user 150 is not obtained, the modulated attack audio may be generated by modulating the initial attack audio.

In step S160 , the hostile user 150 may output the modulated attack audio (eg, the first attack audio 101 of FIG. 1A ).

As described above, the hostile user 150 may perform the voice attack generation step 100, repeating steps S120 to S160 until the completed attack audio is obtained.

If the server 2000 does not detect the voice attack generating step 100 , the voice attack is completed, and the completed voice attack may be hidden in an audio file or a video file. When the hidden voice attack is input to the general user's client device, the hidden voice attack is recognized as the general user's voice command, so that the general user's client device and peripheral devices can be controlled.

The server 2000 according to an embodiment may detect the voice attack generation step 100 . Specifically, the server 2000 may determine whether the input audio signal input in the voice attack generation step 100 is an audio signal for an attack on the voice assistant service. When it is determined that the input audio signal is the input audio signal of the step of generating a voice attack input by the hostile user 150 , the server 2000 may perform an operation for preventing the voice attack generation. A specific method for the server 2000 to detect the step of generating a voice attack will be described in detail with reference to FIGS. 3 to 8 .

2 is a block diagram illustrating a configuration of a server according to an embodiment.

Referring to FIG. 2 , the server 2000 may include a communication interface 2100 , a processor 2200 , and a storage unit 2300 .

The communication interface 2100 may perform data communication with the client device 1000 under the control of the processor 2200 . Also, the communication interface 2100 may perform data communication not only with the client device 1000 , but also with other peripheral devices 3000 .

Communication interface 2100 is, for example, wired LAN, wireless LAN (Wireless LAN), Wi-Fi (Wi-Fi), Bluetooth (Bluetooth), Zigbee (zigbee), WFD (Wi-Fi Direct), infrared communication (IrDA, Infrared Data Association), BLE (Bluetooth Low Energy), NFC (Near Field Communication), WiBro (Wireless Broadband Internet, Wibro), WiMAX (World Interoperability for Microwave Access, WiMAX), SWAP (Shared Wireless Access Protocol), WiGig Data communication may be performed with the client device 1000 or other peripheral devices 3000 using at least one of a data communication method including (Wireless Gigabit Alliance, WiGig) and RF communication.

In order to control the client device 1000 or the peripheral devices 3000 , the communication interface 2100 according to an embodiment receives a voice command and a response request to the voice command from the client device 1000 , and responds to the voice command. A response may be transmitted to the client device 1000 .

The processor 2200 may execute one or more instructions of a program stored in the storage 2300 . The processor 2200 may include hardware components that perform arithmetic, logic, input/output operations and signal processing.

The processor 2200 may include, for example, a central processing unit (Central Processing Unit), a microprocessor, a graphic processor (Graphic Processing Unit), Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), and Digital Signal Processors (DSPDs). Signal Processing Devices), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), Application Processors (APs), Neural Processing Units, or artificial intelligence-only processors designed with a hardware structure specialized for processing artificial intelligence models. It may consist of at least one, but is not limited thereto.

The processor 2200 according to an embodiment may convert the audio signal into text by performing Automated Speech Recognition (ASR) on the audio signal representing the received voice command. Also, the processor 2200 may analyze texts converted from the received audio signal to determine whether the received audio signal is an audio signal input by a user. In addition, the processor 2200 analyzes the text converted from the received audio signal using a natural language understanding model (NLU), and based on the analysis result, the client device 1000 or the peripheral devices 3000 . may generate control information for controlling the , and transmit it to the client device 1000 .

The storage unit 2300 is, for example, a flash memory type, a hard disk type, a multimedia card micro type, or a card type memory (eg, SD or XD). Non-volatile memory including at least one of memory, etc.), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk and volatile memory such as random access memory (RAM) or static random access memory (SRAM).

The storage unit 2300 may store instructions, data structures, and program codes readable by the processor 2200 . In the disclosed embodiments, operations performed by the processor 2200 may be implemented by executing instructions or codes of a program stored in the storage 2300 . In addition, data and program command codes corresponding to the attack management module 2310 , the voice assistant module 2320 , and the database 2330 may be stored in the storage 2300 .

In an embodiment, the processor 2200 determines whether the input audio signal is an audio signal for an attack on the voice assistant service using the data and the command code related to the attack management module 2310, and determines whether the input audio signal is an audio signal for an attack on the voice assistant service. When the signal corresponds to an audio signal for an attack on the voice assistant service, an operation for preventing the generation of a voice attack may be performed.

The attack management module 2310 according to an embodiment may include an audio comparison module 2311 , a text comparison module 2312 , an attack determination module 2313 , and an attack response module 2314 .

The processor 2200 compares the input audio signal with the user's audio signals pre-stored in the database 2330 using the audio comparison module 2311, and compares the input audio signal with the user's audio signal similar to the input audio signal among the pre-stored audio signals. Similar audio signals may be identified.

The processor 2200 compares the text converted from the input audio signal with texts converted from a plurality of audio signals of the user stored in the database 2330 using the audio-text difference calculation module 2312 to compare the text The difference values between them can be calculated.

The processor 2200 uses the attack determination module 2313 to determine whether the input audio signal is an audio signal for an attack on the voice assistant service based on difference values between texts calculated from the text comparison module 2312 can do. Specifically, as similar audio signals are input a predetermined number of times or more during a predetermined period, a predetermined number or more of similar audio signals identified by the audio comparison module 2311 may exist. Also, since there is a difference in text between the identified similar audio signals, the degree of change in text converted from the similar audio signals may be greater than or equal to a predetermined value. The processor 2200 uses the attack determination module 2312 to input a similar audio signal a predetermined number of times or more, and when a variation of texts converted from the similar audio signals is identified as a predetermined value or more, the hostile user 150 ), it can be determined that the voice attack generation step 100 is in progress. In this case, the processor 2200 may determine the current input audio signal as an audio signal included in the voice attack generation step 100 by the hostile user 150 .

When it is determined that the input audio signal is an audio signal for an attack on the voice assistant service, the processor 2200 may perform an operation for preventing the generation of a voice attack using the attack response module 2314 .

Specific operations of using each module included in the attack management module 2310 in order for the processor 2200 to detect and respond to a voice attack generation step will be described later.

In an embodiment, the processor 2200 converts an input audio signal into text using data and command codes related to the voice assistant module 2320 , and analyzes the converted text to control the device using data and command codes can create In this case, the controlled device may be the client device 1000 and the peripheral devices 3000 .

The voice assistant module 2321 may include an ASR model 2321 , an NLU model 2322 , an NLG model 2323 , and a device control information generating module 2324 .

The processor 2200 may perform ASR using the Automatic Speech Recognition (ASR) model 2321 and may convert the audio signal received from the client device 1000 into input text. The ASR model 2321 may include an acoustic model (AM) or a language model (LM), but is not limited thereto.

The processor 2200 may obtain a semantic element in the text by using the NLU model 2322 to interpret the text obtained through the ASR model.

The processor 2200 may generate and edit speech data related to functions of the client device 1000 and the peripheral devices 3000 by using the NLG model 2323 to generate a natural language for conversation with the user.

The processor 2200 uses the device control information generation module 2324 to generate control information for controlling the client device 1000 and peripheral devices based on semantic elements in the text obtained from the NLU model 2322. can

The database 2330 included in the storage unit 2300 according to an embodiment is a user audio signal database 2331 in which the user's audio signals are stored or the user's audio signals are stored in the form of data converted using various pre-processing methods. ), and a user text database 2332 in which texts converted from the user's audio are stored. Data stored in each database may be stored in the form of at least one of an audio signal, data converted from the audio signal, and text converted from the audio signal when a user's audio signal is input.

3 is a flowchart illustrating methods in which a server operates to detect a voice attack on a device according to an embodiment.

In the description of FIG. 3 , the same components as in FIGS. 1A, 1B and 2 will be described using the same reference numerals.

In an embodiment, the input audio signal may be a voice command signal uttered by a general user who wants to use the voice assistant service and received by the client device 1000 . In addition, the input audio signal may be an audio signal recognized as a voice command by the client device 1000 by outputting an adversarial audio generated by the hostile user 150 who wants to generate a voice attack through a speaker or the like. The server 2000 according to the disclosed embodiment may distinguish each user when there are a plurality of users, but before determining whether the input audio signal is an audio signal for a voice attack, whether the user is a general user or a hostile user Since it is not known whether it is (150), the input audio signal will be referred to as the user's input audio signal hereinafter.

In operation S310 , the server 2000 may receive the user's input audio signal from the client device 1000 . The input audio signal may include a wake up command for calling the client device 1000 and a control command for controlling at least one of the client device 1000 and the peripheral devices 3000 , but is limited thereto. it is not going to be

The server 2000 according to an embodiment may receive an input audio signal and perform ASR for converting the received audio signal into computer-readable text.

In operation S320, the server 2000 according to an embodiment may compare the received input audio signal with each of a plurality of audio signals previously stored in the database 2330 within the server.

The plurality of audio signals previously stored in the database 2330 may be generated from audio inputs by a user of the client device 1000 . Also, when there are a plurality of users, a plurality of pre-stored audio signals may be stored separately for each user. When a plurality of pre-stored audio signals for each user are separately stored in the server 2000, the input audio signal is compared with the user's pre-stored audio signals corresponding to the input audio signal among the plurality of pre-stored audio signals . For example, when the input audio signal is the audio signal of the hostile user 150 , the plurality of pre-stored audio signals compared with the input audio signal may be signals obtained by storing the audio signals of the hostile user 150 .

Also, when a plurality of pre-stored audio signals are stored, the amount of stored audio signals may be adjusted using an adjustment parameter. The adjustment parameter is a parameter for efficiently managing data of the database 2330 and may include a period for storing an audio signal and a data type for storing the audio signal, but is not limited thereto.

For example, the plurality of audio signals pre-stored in the database 2330 may be audio signals commanded by the user to the client device 1000 for a period of one week. In addition, when there are a plurality of users of the client device 1000, the plurality of audio signals pre-stored in the database 2330 may be stored separately for each user by applying different adjustment parameters to each user. . Also, the plurality of audio signals pre-stored in the database 2330 may be stored in a converted data form instead of an original form by applying a predetermined algorithm (eg, normalization).

The server 2000 according to an embodiment may store the received input audio signal in a plurality of pre-stored in the database 2330 in order to determine whether audio signals similar to the received input audio signal are input more than a predetermined number of times during a predetermined period. The similarity between the input audio signal and the plurality of pre-stored audio signals may be calculated by comparing the respective audio signals. A detailed method by which the server 2000 compares the received input audio signal with a plurality of audio signals previously stored in the database 2330 will be described in detail with reference to FIGS. 4A and 4B .

In step S330, the server 2000 according to an embodiment may identify audio signals similar to the input audio signal from among the plurality of pre-stored audio signals based on the comparison result of the input audio signal and the plurality of pre-stored audio signals. have. Specifically, as a result of comparing the input audio signal with the plurality of pre-stored audio signals, the server 2000 identifies audio signals having a similarity greater than or equal to a predetermined threshold among the plurality of pre-stored audio signals as similar audio signals. can

The identified similar audio signals may mean that audio signals similar to the input audio signal are received and stored by the server 2000 through the client device 1000 before the input audio signal is input. Accordingly, the server 2000 may identify how many audio signals similar to the input audio signal are input and stored in the server 2000 before the input audio signal is received.

In addition, the threshold, which is a criterion for which the server 2000 according to an embodiment identifies similar audio signals from among a plurality of pre-stored audio signals, includes the number of similar audio signals, a period in which the plurality of pre-stored audio signals are stored, and a plurality of pre-stored audio signals. of the audio signals may be changed based on at least one of the stored data types.

Also, when the number of similar audio signals having a similarity with the input audio signal equal to or greater than a predetermined threshold among the plurality of pre-stored audio signals is greater than the reference value N, the server 2000 determines the similarity to the input audio signal from among the identified similar audio signals. The N signals having the highest n may be identified as similar audio signals.

In operation S340, the server 2000 according to an embodiment may perform automatic speech recognition (ASR) to convert the received input audio signal into the first text. Specifically, the server 2000 may perform ASR for converting an audio signal into computer-readable text using a predefined model such as an acoustic model (AM) or a language model (LM). . Also, when the server 2000 receives an acoustic signal from which noise is not removed from the client device, the server 2000 may obtain an audio signal by removing noise from the received acoustic signal and perform ASR on the audio signal.

Also, the server 2000 according to an embodiment may obtain the converted second texts from the similar audio signals.

In an embodiment, the second texts converted from the analogous audio signals are obtained by the server 2000 performing ASR on the analogous audio signals identified in step S330 and converting the identified analogous audio signals into texts. it may have been

In another embodiment, when a plurality of pre-stored audio signals are stored in the database 2330 , ASR is performed on the audio signals stored by the server 2000 to generate third texts converted from the plurality of audio signals. can be stored together. In this case, the second texts converted from the analogous audio signals may be obtained by the server 2000 selecting texts corresponding to the identified analogous audio signals from among the third texts.

When the first text converted from the input audio signal and the second text converted from the similar audio signals are obtained, the server 2000 according to an embodiment may perform step S350.

In step S350, the server 2000 according to an embodiment converts the text converted from the input audio signal into the analogous audio signal to determine whether the text converted from the audio signals is transformed with respect to the input audio signal and the analogous audio signals. It can be compared with the texts converted from the signals.

In an embodiment, the server 2000 may compare the first text converted from the received input audio signal with the second text converted from the similar audio signals. Specifically, the server 2000 may calculate a similarity between the first text and the second texts by comparing the first text with each of the texts included in the second texts. For example, the server 2000 may generate difference values indicating a difference between the first text and the second texts based on at least one of a difference between characters of the first text and the second text, a difference between words, and a difference between pronunciation expressions. The similarity can be calculated by calculating.

A detailed method for the server 2000 to compare the first text and the second text will be described in detail with reference to FIG. 5 .

In step S360, the server 2000 according to an embodiment may determine whether the input audio signal is an audio signal for attacking the voice assistant service based on a result of comparing the first text and the second text . For example, similar audio signals are identified from among the pre-stored audio signals so that a predetermined number or more of similar audio signals similar to the input audio signal exist. may be different from In this case, as described above with reference to FIGS. 1A and 1B , the input audio signal is a process for confirming texts converted from the similar audio signals while the hostile user 150 modulates and inputs the similar audio signals, a voice attack generation step It may mean that it is an audio signal included in .

The server 2000 according to an embodiment determines that the user's input audio signal is audio for attack on the voice assistant service based on the degree to which the text converted from the input audio signal is different from the text converted from the pre-stored similar audio signals. signal can be determined. That is, the server 2000 may determine whether the user who has input the input audio signal is the hostile user 150 .

Specifically, the server 2000 identifies values in which the difference value between the first text and the second text is equal to or greater than the first threshold value among the difference values indicating the difference between the first text and the second text, and the difference value is equal to or greater than the first threshold value. When the number of values equal to or greater than the first threshold value is equal to or greater than the second threshold value, the input audio signal may be determined as an audio signal for attacking the voice assistant service.

Also, the first threshold value and the second threshold value may be set according to the number of identified similar audio signals. For example, when the number of identified similar audio signals is greater than the predetermined reference value A, a plurality of values representing differences between texts greater than a first threshold value are identified, so that the efficiency of calculation may be reduced. In this case, the first threshold value may be set to a value higher than the current value. In another example, when the identified similar audio signals are less than the predetermined reference value B, the number of values having a difference value equal to or greater than the first threshold value is the second despite the existence of values indicating the difference between texts that are greater than the first threshold value. may be less than the threshold. In this case, for quick determination, the second threshold value may be set to a value lower than the current value.

In addition, when determining whether the input audio signal is an audio signal for attacking the voice assistant service based on the difference between the texts, the server 2000 may further use the comparison result between the input audio signal and a plurality of pre-stored audio signals. can

Also, when determining whether the input audio signal is an audio signal for attacking the voice assistant service based on the difference between the texts, the server 2000 may further use difference values between the second texts. For example, the server 2000 compares the second texts converted from the identified analogous audio signals and calculates difference values between the second texts to further determine the degree to which texts converted from the analogous audio signals are transformed. It can be used to determine whether the input audio signal is an audio signal for an attack on the voice assistant service.

When it is determined that the input audio signal is an audio signal for an attack on the voice assistant service, the server 2000 according to an embodiment may determine a corresponding action for preventing the generation of a voice attack. This will be described in detail in the description of FIG. 7 .

4A and 4B are diagrams for explaining a method in which a server compares an input audio signal with a plurality of pre-stored audio signals using an audio comparison module according to an exemplary embodiment.

In the description of FIGS. 4A and 4B, the same components as in FIGS. 1A, 1B and 2 will be described using the same reference numerals.

The server 2000 according to an embodiment may include an audio comparison module 2311 in the attack management module 2310 . Also, the audio comparison module 2311 may include an audio feature extractor 420 and an audio similarity calculator 430 .

In an embodiment, the audio comparison module 2311 may compare the input audio signal 410 with a plurality of audio signals previously stored in the user audio signal database 2331 .

Referring to FIG. 4A , the input audio signal 410 obtained by the server 2000 may be input to the audio comparison module 2311 .

In one embodiment, the input audio signal 410 is a server as the attack audio generated by the hostile user 150 output from the speaker is recognized by the client device 1000 as an audio signal representing a voice command. may have been sent to

Specifically, when the attack audio of the hostile user 150 includes a voice command for calling the client device 1000 (eg, "Hi Bixby" or "Hi Bixby, open the front door", etc.), The client device 1000 may recognize the attack audio of the hostile user 150 including the call command as the input audio signal 410 and transmit it to the server 2000 . The server 2000 may input the received input audio signal 410 to the audio comparison module 2311 .

When the input audio signal 410 is input, the audio comparison module 2311 according to an embodiment may load a plurality of audio signals previously stored in the user audio signal database 2331 . The plurality of audio signals pre-stored in the user audio signal database 2331 may be audio signals input by the user for a predetermined period of time. Also, when there are a plurality of users, a plurality of pre-stored audio signals may be stored separately for each of the plurality of users. When a plurality of pre-stored audio signals for each user are separately stored in the server 2000, the input audio signal is compared with the pre-stored audio signals of the user who input the input audio signal among the plurality of pre-stored audio signals . For example, when the input audio signal is an audio signal of the hostile user 150 , the input audio signal may be compared with a plurality of pre-stored audio signals of the hostile user 150 .

The audio comparison module 2311 may compare the input audio signal with each of a plurality of pre-stored audio signals. In this case, the input audio signal and the plurality of pre-stored audio signals may be analog audio signals, and comparison between the input audio signal and the plurality of pre-stored audio signals may be performed in a time domain or a frequency domain.

The audio comparison module 2311 may compare the waveform of the input audio signal with the waveform of each of the plurality of pre-stored audio signals using the audio similarity calculator 430 . For example, the audio comparison module 2311 converts an input audio signal and a plurality of pre-stored audio signals into a digital signal, performs sampling and quantization, and then compares the waveform of the input audio signal and the plurality of pre-stored audio signals. The similarity between the waveforms can be calculated.

Also, the audio comparison module 2311 may perform a preprocessing process on the input audio signal using the audio feature extractor 420 and extract feature vectors from the input audio signal and a plurality of pre-stored audio signals. The audio comparison module 2311 may calculate the similarity between the input audio signal and a plurality of pre-stored audio signals by comparing the extracted feature vectors using the audio similarity calculator 430 .

In an embodiment, various methods may be used for the method for the audio feature extractor 420 and the audio similarity calculator 430 to extract and compare feature vectors from an audio signal. For example, methods such as correlation coefficient analysis, cepstrum analysis, and MFCC coefficient (Mel Frequency Cepstral Coefficient) analysis may be used, but the present invention is not limited thereto.

The audio comparison module 2311 according to an embodiment compares the input audio signal and each of a plurality of audio signals previously stored in the user audio signal database 2331, and calculates the similarity, which is difference values indicating the difference between the audio signals, Among the plurality of pre-stored audio signals, audio signals having a difference value greater than or equal to a predetermined threshold may be identified as similar audio signals.

Also, the audio comparison module 2311 may generate the similar audio signal set 440 by grouping the identified similar audio signals. In this case, if the number of identified similar audio signals is greater than the predetermined number of N, the audio comparison module 2311 may select N similar audio signals having the highest similarity to generate the similar audio signal set 440 . The audio comparison module 2311 may output the generated similar audio signal set 440 and the input audio signal 410 .

In another embodiment, for efficiency of operation, the audio comparison module 2311 converts audio signals into different data types and compares the converted data to calculate the similarity between the input audio signal and a plurality of pre-stored audio signals. .

Referring to FIG. 4B , a plurality of audio data may be stored in the user audio signal database 2331 . In this case, the plurality of pre-stored audio data may be data converted from the plurality of audio signals representing the user audio signal.

For example, when a plurality of audio signals are stored in the user audio signal database 2331 of the plurality of pre-stored audio data, a predetermined data processing method is applied to the plurality of audio signals to convert the converted user audio data into the user audio data. It may be stored with the signals.

In another example, the plurality of pre-stored audio data does not store the plurality of audio signals when the plurality of audio signals are input, but user audio data converted by applying a predetermined data processing method to the plurality of audio signals It may be stored instead of audio signals.

The plurality of pre-stored audio data according to an embodiment may be data in a converted form using at least one method of normalization, locality-preserving hashing (LPH), and clustering. . However, the data types of the plurality of pre-stored audio data are not limited to the above-described examples, and the plurality of pre-stored audio data may be stored in various data types in order to reduce the complexity of the operation.

The input audio data 450 according to an embodiment may be input to the audio comparison module 2311 . In this case, the predetermined data processing method applied to the input audio signal 410 may be a method corresponding to the data processing method applied to the plurality of audio data previously stored in the user audio signal database 2331 .

The audio comparison module 2311 may identify similar audio data by comparing the input audio data with each of a plurality of pre-stored audio data using the audio similarity calculator 430 . Also, when there are a plurality of users, a plurality of pre-stored audio data may be stored separately for each of the plurality of users. When a plurality of pre-stored audio data for each user are separately stored in the server 2000, the input audio data is compared with the pre-stored audio data of the user who input the input audio data among the plurality of pre-stored audio signals. . For example, when the input audio signal is an audio signal of the hostile user 150 , the input audio signal may be compared with a plurality of pre-stored audio signals of the hostile user 150 .

In an embodiment, the input audio data 450 and the plurality of pre-stored audio data may be normalized waveform data. The audio comparison module 2311 applies a peak alignment algorithm for matching the peak values of the normalized waveform data, and measures the cross-correlation of the adjusted waveforms to match the peak values and inputs A similarity between the audio data 450 and a plurality of pre-stored audio data may be calculated. The audio comparison module 2311 may identify similar audio data similar to the input audio data based on the similarity calculation result.

In another embodiment, the input audio data 450 and the plurality of pre-stored audio data may be hash values to which a locality-preserving hashing (LPH) algorithm is applied. The audio comparison module 2311 may calculate a similarity between the input audio data 450 and the plurality of pre-stored audio data by comparing a hash value representing the input audio data 450 with hash values representing a plurality of pre-stored audio data. . The audio comparison module 2311 may identify similar audio data similar to the input audio data based on the similarity calculation result.

In another embodiment, the input audio data 450 and the plurality of pre-stored audio data may be data in which a clustering algorithm is applied to cluster similar entities. In this case, since objects within the same cluster are classified into the same cluster after the similarity is calculated, the audio comparison module 2311 applies a clustering algorithm to the input audio data 450 and within the cluster to which the input data 450 is assigned. Similar audio data can be identified by identifying the entities.

The audio comparison module 2311 according to an embodiment may generate the similar audio data set 460 by grouping the identified similar audio data. In this case, if the number of identified similar audio data is greater than the predetermined number N, the audio comparison module 2311 may select N similar audio data having the highest similarity to generate the similar audio data set 460 . The audio comparison module 2311 may output the generated similar audio data set 460 and input audio data 450 . However, the present invention is not limited thereto, and the audio comparison module 2311 outputs the similar audio signal set 440 corresponding to the generated similar audio data set 460 and the input audio signal 410 corresponding to the input audio data together. You may.

FIG. 5 is a diagram for describing a method in which a server compares first text converted from an input audio signal with second text converted from a plurality of pre-stored audio signals using a text comparison module, according to an exemplary embodiment.

In the description of FIG. 5 , the same components as in FIGS. 1A, 1B, 4A and 4B will be described using the same reference numerals.

The server 2000 according to an embodiment may include a text comparison module 2312 in the attack management module 2310 . Also, the text comparison module 2312 may include a text acquisition unit 510 and a text comparison unit 520 .

Referring to FIG. 5 , an input audio signal 410 and a similar audio signal set 440 may be input to a text comparison module 2312 . However, the present invention is not limited thereto, and the input audio data 450 and the similar audio data set 460 illustrated in FIG. 4B may be input to the text comparison module 2312 .

The text comparison module 2312 according to an embodiment may include a first text 501 converted from an input audio signal and a second text converted from similar audio signals included in a similar audio signal set using the text obtaining unit 510 . 2 texts 502 may be obtained.

Specifically, the first text 501 converted from the input audio signal may be obtained by inputting the input audio signal to the ASR model 2321 stored in the server, and performing ASR converting the input audio signal into text. . In this case, the text obtaining unit 510 may request to perform ASR on the input audio signal and obtain the converted first text 501 from the input audio signal.

In addition, the second texts 502 converted from the analogous audio signals included in the analogous audio signal collection are input to the ASR model 2321 stored in the server, and the analogous audio signals are converted into texts. These may be those obtained by performing ASR. In this case, the text obtaining unit 510 may request to perform ASR on the analogous audio signals and obtain the converted second texts 502 from the similar audio signals.

Also, although not shown in FIG. 5 , the second texts 502 obtained by the text obtaining unit 510 may be obtained from third texts previously stored in the user text database 2332 . Specifically, when a plurality of audio signals are stored in the user audio signal database 2331, ASR is performed on the plurality of audio signals to be converted into text, and third texts converted from the plurality of audio signals are stored in the user text database (2332) may be stored. In this case, the text acquisition unit 510 may acquire the second texts 502 that are texts corresponding to the similar audio signals from among the third texts previously stored in the user text database 2332 .

The text comparison module 2312 according to an embodiment includes the first text 501 converted from the input audio signal and the second text converted from the analogous audio signals included in the analogous audio signal set using the text comparison unit 520 . The two texts 502 can be compared.

For example, the text comparison module 2312 may compare the first text 501 with each of the second texts 502 to calculate difference values 530 indicating a difference between the first text and the second texts. have. For example, a first difference value 531 is obtained by comparing the first text and audio similar text included in the second text, and the first text and the second audio text text included in the second text are compared. Thus, a second difference value 532 may be obtained, and a third difference value 533 may be obtained by comparing the first text and the third similar audio text included in the second text.

A method in which the text comparison module 2312 according to an embodiment calculates the difference values 530 representing the difference between the first text and the second text may be performed by applying various algorithms.

For example, the method of calculating the difference values 530 between the first text and the second texts includes characters included in the first text 501 and characters included in each of the second texts 502 . It may be a character-wise edit distance algorithm that calculates difference values based on the difference between characters.

In another example, the method of calculating the difference values 530 between the first text and the second texts includes words included in the first text 501 and words included in each of the second texts 502 . It may be a word-wise edit distance algorithm that calculates difference values 530 representing differences between texts based on the difference between words.

In another example, the method of calculating the difference values 530 between the first text and the second texts is based on the difference between the phonetic representation of the first text 501 and the phonetic representations of each of the second texts 502 . Based on it, it may be a phonetic representation edit distance algorithm that calculates difference values. However, the present invention is not limited to the above-described examples, and the text comparison module 2312 may calculate difference values 530 indicating a difference between the first text and the second text using various methods. Also, the text comparison module 2312 may calculate difference values indicating a difference between the second texts by using the above-described methods.

The text comparison module 2312 according to an embodiment may output difference values 530 indicating a difference between the calculated first text and the second text.

6 is a diagram for explaining a method for a server to determine whether an input audio signal is an audio signal for an attack on a voice assistant service using an attack determination module, according to an embodiment.

In the description of FIG. 6 , the same components as in FIGS. 1A, 1B to 5 will be described using the same reference numerals.

Referring to FIG. 6 , text difference values 530 output from the text comparison module may be input to the attack determination module 2313 .

In an embodiment, the attack determination module 2313 may determine whether the input audio signal is an audio signal for an attack on the voice assistant service, based on the text difference values 530 . For example, as a result of comparing the input audio signal and the pre-stored audio signals, the analogous audio signals are input several times, and text converted from the input audio signal is transformed from the text converted from the written analogous audio signals. ), it may be determined that the input audio signal is an audio signal for attacking the voice assistant service.

Specifically, the attack determination module 2313 may identify text difference values having a text difference value equal to or greater than a first threshold value from among the text difference values 530 comparing the first text and the second texts.

In one embodiment, the values of the text difference value equal to or greater than the first threshold value mean that the input audio signal has a similar waveform to similar audio signals, but texts converted by performing ASR on each audio signal are different it could be Accordingly, when there are values in which the text difference value is equal to or greater than the first threshold value, the input audio signal may be an audio signal by the user (the hostile user 150 ) who intends to generate a voice attack.

Specifically, the input audio signal is input so that the text converted from the input audio signal is converted differently from the texts converted from the previously input similar audio signals in order to find the text corresponding to the voice attack signal desired by the hostile user 150 It may be an audio signal generated by transforming an audio signal.

The attack determination module 2313 may determine the input audio signal as an audio signal for generating a voice attack when values of the text difference value equal to or greater than the first threshold value are equal to or greater than the second threshold value. This means that the number of second texts indicating that the degree of deformation of the first text with respect to the second texts is equal to or greater than the first threshold value and that the degree of deformation of the first text is equal to or greater than the first threshold value is greater than or equal to the second threshold value. can mean doing In this case, the attack determination module determines that the input audio signal is an audio signal for an attack on the voice assistant service, and outputs data indicating that the input audio signal is an audio signal by the hostile user 150 to the attack response module 2314 can do.

When receiving data indicating that the input audio signal is an audio signal by the hostile user 150 , the attack response module 2314 may perform an operation for defending against a voice attack.

In addition, when it is determined that the input audio signal is not an audio signal for generating a voice attack, the attack determination module 2313 transmits data indicating that the input audio signal is not an audio signal for generating a voice attack to the voice assistant module 2320 ) can be printed.

When receiving data indicating that the input audio signal is not an audio signal for generating a voice attack, the voice assistant module 2320 may transmit a response to the input audio signal to the client device.

Specifically, the voice assistant module 2320 may input the text converted from the input audio signal into the NLU model 2322, parse the text, and analyze intents included in the parsed text. In addition, the voice assistant module 2320 generates control information for controlling the client device and peripheral devices by inputting the intent analysis result to the device control information generation module 2324 , and transmits the control information to the client device 1000 . can

In this case, the client device 1000 may receive a response to the input audio signal to control the operation of the client device 1000 or the peripheral devices 3000 .

7 is a diagram for describing a method for a server to determine an operation for preventing generation of a voice attack by using an attack response module according to an embodiment.

In an embodiment, when it is determined that the input audio signal is an audio signal for an attack on the voice assistant service, the attack response module 2314 may perform an operation for defending it.

For example, when the attack response module 2314 receives data indicating that the input audio signal is an audio signal by the hostile user 150 from the attack determination module 2313, the hostile user 150 who requests a response to the voice command ) account can be blocked. The hostile user's account may be a user account of the voice assistant service of the server 2000 , and as the hostile user's account is blocked, provision of the voice assistant service to the hostile user may be stopped. In this case, the server 2000 may not respond to a subsequent input audio signal of the hostile user 150 even when a response to the voice command is requested. Therefore, the hostile user 150 who wants to generate a voice attack transcribes the input audio signals input next to the currently input audio signal into text corresponding to the voice attack desired by the hostile user 150. It is not possible to check whether

In another example, when the attack response module 2314 receives data indicating that the input audio signal is an audio signal for generating a voice attack from the attack determination module 2313, the first text converted from the input audio signal is not returned. , and can return any other text.

For example, the attack response module 2314 may return one text among the second texts converted from the similar audio signals instead of the first text converted from the input audio signal. However, the present invention is not limited thereto, and the attack response module 2314 may return an arbitrary text instead of the first text or may not return any text. Accordingly, the hostile user 150 attempting to generate a voice attack cannot confirm the first text corresponding to the currently inputted audio signal.

8 is a diagram for describing another embodiment in which a server operates to detect generation of a voice attack on a device according to an embodiment.

According to one embodiment, the user may be a hostile user 150 attempting to create a voice attack. In this case, the attack audio of the hostile user 150 may be output by a speaker or the like and transmitted to the client device 1000 as an input in the form of an audio signal.

In the client device 1000 according to an embodiment, when a call command is included in an audio signal input to the client device, a function for providing a voice assistant service is activated in response to the call command and requests a response from the server 2000 . can

In the process of the hostile user 150 generating a voice attack, when the generation of the voice attack is incomplete, the attack audio of the hostile user 150 input to the client device 1000 does not include an accurate call command, so the client The device 1000 may not be activated. However, since the server 2000 needs to detect an event in which similar audio signals are repeatedly input and check and compare texts converted from the similar audio signals to prevent the generation of a voice attack, the server 2000 does not include an accurate call command. It is necessary to store similar audio signals for similar audio signals, and to check texts converted from the similar audio signals. In this case, the client device 1000 may be activated for audio signals including a similar call command rather than an exact call command, and may request a response to the voice command from the server 2000 .

For example, the correct call command for calling the client device 1000 may be “Hi Bixby”, but the attack audio of the hostile user 150 is the exact call command “Hi Bixby (Hi, Bixby)" may not be included.

Specifically, the call command included in the attack audio to the hostile user 150 is a similar call command such as "Hi Bix", "Hi Bisby", "Hi Bibi", etc. can take In this case, the client device 1000 according to an embodiment may be activated by the similar call commands to transmit an audio signal including the similar call command to the server 2000 . The server 2000 according to an embodiment stores an audio signal including a similar call command in an audio signal database to generate a database of pre-stored similar audio signals, or stores an audio signal including a similar call command in a pre-stored similarity audio signal database. By comparing the audio signals, it may be determined whether the audio signal including the similar call command is an audio signal for generating a voice attack. A method for the server to determine whether an input audio signal is an audio signal for generating a voice attack has been described above with reference to FIG. 3 , and thus a description thereof will be omitted.

9 is a diagram for describing a method in which a server receives an input audio signal in a database and generates a plurality of pre-stored audio signals according to an exemplary embodiment.

In operation S910 , the server 2000 may receive and obtain an input audio signal from the client device 1000 . Since this is the same as step S310 of FIG. 3 , a description thereof will be omitted.

In step S920 , the server 2000 may store the input audio signal in the database 2330 . The input audio signal may be stored in the original form, and may be stored in the converted data form by applying a predetermined data processing method to the input audio signal as described with reference to FIG. 4B . Also, when there are a plurality of users, the input audio data may be stored separately for each of the plurality of users.

In operation S930, the server 2000 may check whether the input audio signal is recognized as an accurate voice command. For example, the server 2000 may perform ASR on the input audio signal to convert it into text, and analyze the text using the NLU model. As a result of the text analysis, the server 2000 according to an embodiment provides an accurate voice command for controlling the client device 1000 (eg, “Hi Bixby, tell me today’s schedule”) or an accurate voice command for controlling a peripheral device. (eg, "Hi Bixby, open the front door"), a response to the input audio signal may be transmitted to the client device (step S935).

In operation S940 , the client device 1000 according to an embodiment may receive a response to the voice command from the server 2000 to control the client device 1000 and peripheral devices. For example, when the voice command of a general user who wants to use the voice assistant is “Hi Bixby, tell me about today’s schedule”, the client device 1000 may output an audio briefing the general user’s today’s schedule. In another example, when the general user's voice command is "Hi Bixby, open the front door," the client device 1000 may control the peripheral device 3000 attached to the front door to open the front door.

Referring back to step S930, the server 2000 may check whether the input audio signal is recognized as an accurate voice command. For example, the server 2000 may perform ASR on the input audio signal to convert it into text, and analyze the text using the NLU model. As a result of text analysis, when the meaning of the text is unclear and thus cannot be recognized as a voice command, the input audio signal may be an inaccurately uttered audio signal or an audio signal for generating a voice attack. When the input audio signal is not recognized as an accurate voice command, the server 2000 may perform step S950.

In operation S950 , the server 2000 may check whether a plurality of audio signals previously stored in the database 2330 exist.

When a plurality of audio signals pre-stored in the database 2330 exist, the server 2000 may perform operations of determining whether the input audio signal is an audio signal for generating a voice attack.

When the plurality of audio signals previously stored in the database 2330 do not exist, the server 2000 may end the operation.

10 is a flowchart illustrating a method in which a server detects generation of a voice attack and performs a device control operation or a defense operation of voice attack generation based on a detection result, according to an exemplary embodiment.

In steps S1010 to S1060, the server 2000 may receive an input audio signal, and detect whether the received input audio signal is an audio signal for generating a voice attack. Since this corresponds to steps S310 to S360 of FIG. 3 , a description thereof will be omitted herein.

When it is determined that the received input audio signal is not the audio signal of the voice attack generation step, the server 2000 may transmit a response to the input audio signal to the client device 1000 (step S1065).

In operation S1070 , the client device 1000 may receive a response to the input audio signal to control the operation of the client device 1000 or the peripheral devices 3000 . Since this corresponds to step S940 of FIG. 9 , a description thereof will be omitted herein.

When it is determined that the received input audio signal is the audio signal of the voice attack generation step, the server 2000 may perform step S1080.

In step S1080, it is possible to determine the operation of the server to defend the voice attack generation step. For example, a text other than the text converted from the input audio signal may be returned, or a user determined to be generating a voice attack may be blocked. Since this has been described above in the description of FIG. 7 , a description thereof will be omitted herein.

11 is a block diagram of a client device according to an embodiment.

Referring to FIG. 11 , the client device 1000 according to an embodiment may include an input unit 1100 , an output unit 1200 , a processor 1300 , a memory 1400 , and a communication interface 1500 .

According to an embodiment, there may be at least one client device 1000 . Also, the client device 1000 may be, for example, a smart phone, a tablet PC, a PC, a laptop, a smart TV, a smart refrigerator, or an artificial intelligence speaker in which a voice assistant may be mounted, but is not limited thereto.

The input unit 1100 means a means for a user to input data for controlling the client device 1000 . For example, the input unit 1100 includes a microphone, a key pad, a dome switch, and a touch pad (contact capacitive method, pressure resistance film method, infrared sensing method, surface ultrasonic conduction method, red There may be a mechanical tension measurement method, a piezo effect method, etc.), a jog wheel, a jog switch, and the like, but is not limited thereto.

The input unit 1100 may receive a user's voice command for calling the client device 1000 .

The output unit 1200 may output an audio signal, a video signal, or a vibration signal, and the output unit 1200 may include at least one of a display unit, a sound output unit, and a vibration motor.

The processor 1300 controls the overall operation of the client device 1000 . For example, the processor 1300 may generally control the input unit 1100 , the output unit 1200 , the memory 1400 , the communication interface 1500 , and the like by executing programs stored in the memory 1400 . .

The processor 1300 may receive the input audio signal, recognize a call command included in the input audio signal, and transmit a response request to the voice command to the server 2000 . In addition, the processor 1300 may receive and output a query message for generating and editing utterance data related to functions of the client device 1000 and peripheral devices from the server 2000 .

The communication interface 1500 may include one or more components that allow communication with the server 2000 , the IoT cloud server (not shown), and a peripheral device (not shown). For example, the communication interface 1500 may include a short-range communication unit and a mobile communication unit.

Short-range wireless communication unit, Bluetooth communication unit, BLE (Bluetooth Low Energy) communication unit, near field communication unit (Near Field Communication unit), WLAN (Wi-Fi) communication unit, Zigbee communication unit, infrared (IrDA, infrared) Data Association) communication unit, WFD (Wi-Fi Direct) communication unit, UWB (ultra wideband) communication unit, and may include an Ant+ communication unit, but is not limited thereto.

The mobile communication unit transmits/receives a radio signal to and from at least one of a base station, an external terminal, and a server on a mobile communication network. Here, the wireless signal may include various types of data according to transmission and reception of a voice signal, a video call signal, or a text/multimedia message.

The memory 1400 may store a program for processing and control of the processor 1300 , and may store data input to or output from the client device 1000 .

The memory 1400 may include a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (eg, SD or XD memory), and a RAM. (RAM, Random Access Memory) SRAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk , may include at least one type of storage medium among optical disks.

The SDK module (not shown) may be executed by the processor 1300 to perform operations necessary for controlling the client device 1000 and peripheral devices (not shown). The SDK module (not shown) may be downloaded from the server 2000 and installed in the client device 1000 . The SDK module (not shown) may output a GUI for controlling the client device 1000 and peripheral devices (not shown) on the screen of the client device 1000 . If the client device 1000 is an apparatus that does not include a display device, the SDK module (not shown) provides a voice message for the client device 1000 to control the client device 1000 and peripheral devices (not shown). can be output. The SDK module (not shown) may allow the client device 1000 to receive a response from a user and provide it to the server 2000 .

Meanwhile, block diagrams of the client device 1000 and the server 2000 illustrated in FIGS. 2 and 11 are block diagrams for an exemplary embodiment. Each component in the block diagram may be integrated, added, or omitted according to the specifications of each device that is actually implemented. That is, two or more components may be combined into one component, or one component may be subdivided into two or more components as needed. In addition, the function performed in each block is for describing the embodiments, and the specific operation or device does not limit the scope of the present invention.

The method of operating a server according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

In addition, the method of operating a server according to the disclosed embodiments may be provided by being included in a computer program product. Computer program products may be traded between sellers and buyers as commodities.

The computer program product may include a S/W program and a computer-readable storage medium in which the S/W program is stored. For example, computer program products may include products (eg, downloadable apps) in the form of S/W programs distributed electronically through manufacturers of electronic devices or electronic markets (eg, Google Play Store, App Store). have. For electronic distribution, at least a portion of the S/W program may be stored in a storage medium or may be temporarily generated. In this case, the storage medium may be a server of a manufacturer, a server of an electronic market, or a storage medium of a relay server temporarily storing a SW program.

The computer program product, in a system consisting of a server and a client device, may include a storage medium of the server or a storage medium of the client device. Alternatively, if there is a third device (eg, a smart phone) that is communicatively connected to the server or the client device, the computer program product may include a storage medium of the third device. Alternatively, the computer program product may include the S/W program itself transmitted from the server to the client device or the third device, or transmitted from the third device to the client device.

In this case, one of the server, the client device and the third device may execute the computer program product to perform the method according to the disclosed embodiments. Alternatively, two or more of a server, a client device, and a third device may execute a computer program product to distribute the method according to the disclosed embodiments.

For example, the server may execute a computer program product stored in the server to control a client device communicatively connected with the server to perform the method according to the disclosed embodiments.

Although the embodiments have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concept of the present invention as defined in the following claims are also included in the scope of the present invention. belongs to

Claims (20)

A method for a server providing a voice assistant service to detect a voice attack on a device, the method comprising:
receiving an input audio signal from the device;
comparing the received input audio signal with a plurality of audio signals received from the device and stored in the server;
identifying similar audio signals of the input audio signal from among the plurality of pre-stored audio signals based on a comparison result of the received input audio signal and the plurality of pre-stored audio signals;
converting the received input audio signal into first text by performing Automatic Speech Recognition (ASR);
obtaining transformed second texts from the identified similar audio signals;
comparing the first text converted from the input audio signal with the second text converted from the similar audio signals; and
determining whether the input audio signal is an audio signal for attacking the voice assistant service based on a comparison result of the first text and the second text;
includes,
wherein the plurality of pre-stored audio signals are provided from the device to the server before the input audio signal is received. According to claim 1,
Comparing the received input audio signal with the plurality of pre-stored audio signals comprises:
converting the received input audio signal and the plurality of pre-stored audio signals into an audio data format;
comparing audio data converted from the received input audio signal and a plurality of audio data converted from the plurality of pre-stored audio signals;
including,
The audio data converted from the received audio signal and the plurality of audio data converted from the plurality of pre-stored audio signals are,
Normalization (Normalization), locality-preserving hashing (Locality-preserving hashing; LPH), the method that is transformed using at least one method of clustering (Clustering). According to claim 1,
The step of identifying the similar audio signals comprises:
identifying, among the pre-stored audio signals, audio signals having a similarity with the input audio signal equal to or greater than a predetermined threshold as the similar audio signals; and
and generating a similar audio signal set by grouping N similar audio signals having the highest similarity to the input audio signal from among the identified similar audio signals. 4. The method of claim 3,
Obtaining converted second texts from the identified similar audio signals comprises:
The ASR is performed to convert the analogous audio signals included in the analogous audio signal set into texts to obtain the second texts. 4. The method of claim 3,
Obtaining converted second texts from the identified similar audio signals comprises:
and obtaining, as the second texts, texts corresponding to the analogous audio signals included in the identified analogous audio signal set from among the third texts pre-stored in the server. 6. The method of claim 5,
Comparing the texts comprises:
calculating difference values representing a difference between the first text and the second text;
The step of determining whether the input audio signal is an audio signal for an attack on the voice assistant service comprises:
Determining whether an audio signal for an attack on the voice assistant service is based on the calculated difference values between the texts. 7. The method of claim 6,
Comparing the texts comprises:
calculating difference values representing the difference between the first text and the second text by applying an edit distance algorithm based on at least one of a difference between characters, a difference between words, and a difference between pronunciation expressions How to. 7. The method of claim 6,
The step of determining whether the input audio signal is an audio signal for an attack on the voice assistant service comprises:
Among the difference values indicating the difference between the first text and the second texts, when the number of values having a difference value equal to or greater than the first threshold value is equal to or greater than the second threshold value, the input audio signal is converted to an audio for attack on the voice assistant service A method comprising determining with a signal. 9. The method of claim 8,
and the first threshold value and the second threshold value are set according to the number of the identified similar audio signals. 7. The method of claim 6,
The step of determining whether the input audio signal is an audio signal for an attack on the voice assistant service comprises:
and determining the audio signal for an attack on the voice assistant service by further using a comparison result between the received input audio signal and the plurality of pre-stored audio signals. A server providing a voice assistant service for detecting a voice attack on a device, the server comprising:
a communication interface for performing data communication with the device;
a storage unit for storing a program including one or more instructions; and
A processor for executing one or more instructions of the program stored in the storage unit,
The processor is
control the communication interface to receive an input audio signal from the device;
comparing the received input audio signal with a plurality of audio signals received from the device and pre-stored in the storage;
based on a comparison result of the received input audio signal and the plurality of pre-stored audio signals, identify similar audio signals of the input audio signal from among the plurality of pre-stored audio signals;
ASR (Automatic Speech Recognition) is performed to convert the received input audio signal into a first text,
obtaining transformed second texts from the identified similar audio signals;
comparing the first text converted from the input audio signal with the second text converted from the similar audio signals;
Based on the comparison result of the first text and the second text, it is determined whether the input audio signal is an audio signal for attacking the voice assistant service,
The server, wherein the plurality of pre-stored audio signals are provided from the device to the server before the input audio signal is received. 12. The method of claim 11,
The processor is
converting the received input audio signal and the plurality of pre-stored audio signals into an audio data form;
Comparing the audio data converted from the received input audio signal and the plurality of audio data converted from the plurality of pre-stored audio signals,
The audio data converted from the received audio signal and the plurality of audio data converted from the plurality of pre-stored audio signals are,
Normalization (Normalization), locality-preserving hashing (Locality-preserving hashing; LPH), the server that is transformed using at least one method of clustering (Clustering). 12. The method of claim 11,
The processor is
identifying audio signals having a similarity with the input audio signal equal to or greater than a predetermined threshold among the pre-stored audio signals as the similar audio signals;
and generating a similar audio signal set by grouping N similar audio signals having the highest similarity to the input audio signal from among the identified similar audio signals. 14. The method of claim 13,
The processor is
and converting the analogous audio signals included in the analogous audio signal set into texts by performing the ASR to obtain the second texts. 14. The method of claim 13,
The processor is
and obtaining, as the second texts, texts corresponding to the analogous audio signals included in the identified analogous audio signal set from among the third texts previously stored in the storage unit. 16. The method of claim 15,
The processor is
calculating difference values representing a difference between the first text and the second text;
The server determines whether it is an audio signal for an attack on the voice assistant service based on the calculated difference values between the texts. 17. The method of claim 16,
The processor is
A server for calculating difference values indicating a difference between the first text and the second texts by applying an edit distance algorithm based on at least one of a difference between characters, a difference between words, and a difference between pronunciation expressions. 17. The method of claim 16,
The processor is
Among the difference values indicating the difference between the first text and the second texts, when the number of values having a difference value equal to or greater than the first threshold value is equal to or greater than the second threshold value, the input audio signal is converted to an audio for attack on the voice assistant service Judging by the signal, the server. 17. The method of claim 16,
The processor is
The server further determines as an audio signal for attacking the voice assistant service using a comparison result of the received input audio signal and the plurality of pre-stored audio signals. A computer-readable recording medium in which a program for executing the method of claim 1 in a computer is recorded.

Images