KR20240033200A - Pattern learning and abnormal behavior classification methods for EM emission of on-device AI SoC neural network acceleration engines - Google Patents

Abstract

The present inventor's pattern learning and abnormal behavior classification method for electromagnetic wave emissions from the on-device AI SoC neural network acceleration engine includes measuring the electromagnetic wave side channel generated from the power supply device of the on-device AI equipment, and measuring the electromagnetic wave side channel generated from the power supply device of the on-device AI equipment. A step of generating data as a data format file, a step of converting the generated data into a picture format to perform deep learning learning and classification, and examining the correlation between the operating characteristics of the artificial intelligence engine classified by deep learning and abnormal security behavior. Including the step of checking, the checking step may test security vulnerabilities of the corresponding on-device AI by examining the relevance of abnormal security behavior associated with the operation of the artificial intelligence engine.

Description

Pattern learning and abnormal behavior classification methods for EM emission of on-device AI SoC neural network acceleration engines}

The present invention relates to a pattern learning and abnormal behavior classification method for electromagnetic wave emissions from an on-device AI SoC neural network acceleration engine, and more specifically, to the electromagnetic wave side channel of a single chip system (SoC, System on Chip) including an artificial intelligence engine. This is about a system that uses characteristics to learn artificial intelligence engine operation data sets and detect and verify security abnormal behavior for artificial intelligence operation by correlating it with security abnormal behavior signals.

A single chip system (SoC, System on Chip) including an artificial intelligence engine that supports on-device AI is configured to include the artificial intelligence engine function, as shown in Figure 1.

On-device AI is equipment that supports artificial intelligence functions in real time in edge computing equipment terminals, etc., and for this purpose, SoC including an artificial intelligence engine is used.

On-device AI equipment processes important personal information, personal physical information, and social information, so information protection and secure communication are very important. Therefore, it is about a method that can verify security abnormal behavior detection with artificial intelligence operation and a systemized system that can confirm the protection and security of important information in on-device AI.

In the SoC that supports on-device AI, a direct program that runs the artificial intelligence engine that actually performs the artificial intelligence operation is measured and the electromagnetic wave side channel characteristics generated when the operation is performed are converted to side channel information. Construct the dataset. The dataset is learned using deep learning and used as electromagnetic side channel learning data for the artificial intelligence engine.

For on-device AI equipment using the same SoC, the artificial intelligence operation program is run, and the side channel characteristics are measured using electromagnetic wave collection equipment to show the actual operating state of the artificial intelligence engine, and the electromagnetic wave component is used to execute the learned artificial intelligence engine. While comparing with channel information, classification by artificial intelligence engine characteristics is performed using deep learning. By performing correlation analysis between security abnormalities such as abnormal communication, abnormal encryption operation, and key information transmission that occur in the on-device AI equipment being tested, and the main operations of the artificial intelligence engine, the artificial intelligence of the on-device AI is performed. We propose a method and system for detecting and confirming abnormal behavior and performing security tests on on-device AI equipment.

Domestic Patent Publication No. 10-2022-0055378 (Publication date: May 3, 2022)

In order to propose the above system, one task of the present invention is to provide a process for securing a training data set.

Another task of the present invention is to apply deep learning to edge equipment and security camera equipment that use on-device AI to provide a structure and process for an overall system that performs security verification of the AI of the product. .

Another task of the present invention is to provide a process for applying reasoning at the product and service level.

The technical problems to be achieved in the present invention are not limited to the above technical problems, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below.

The pattern learning and abnormal behavior classification method for electromagnetic wave emissions from an on-device AI SoC neural network acceleration engine according to an embodiment of the present invention to solve the above-described problems is an electromagnetic wave side channel generated from the power supply device of on-device AI equipment. A step of measuring, a step of generating data from the measured electromagnetic wave side channel as a radio wave characteristic data format file, a step of converting the generated data into a picture format to perform deep learning learning and classification, and artificial data classified by deep learning. It includes the step of examining the correlation between the operating characteristics of the intelligence engine and abnormal security behavior, wherein the inspection step examines the relevance of abnormal security behavior associated with the operation of the artificial intelligence engine to test security vulnerabilities of the corresponding on-device AI. You can.

According to the present invention as described above, the effects described below can be obtained. However, the effects that can be achieved through the present invention are not limited to this.

First, it has the effect of providing a process for securing a training data set.

Second, by applying deep learning to edge equipment and security camera equipment that use on-device AI, there is an effect of providing the structure and process for the overall system that performs security verification of the AI of the product.

Third, it has the effect of providing a process for applying reasoning at the product and service level.

1 is a diagram of a single chip system including an artificial intelligence engine function according to an embodiment of the present invention.
Figure 2 is a diagram showing a schematic configuration of an on-device AI equipment artificial intelligence engine security abnormal behavior detection and testing method according to an embodiment of the present invention.
Figure 3 is a flowchart configuring a set of electromagnetic wave side channel characteristic learning data for an artificial intelligence engine in reference equipment according to an embodiment of the present invention.
Figure 4 shows various types of on-device AI equipment using the same artificial intelligence engine, based on the operation of an artificial intelligence engine learned based on deep learning according to an embodiment of the present invention, in the test room and field. This is a flowchart that detects abnormal security behavior regarding the artificial intelligence operation of the equipment while operating.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. The detailed description set forth below in conjunction with the accompanying drawings is intended to illustrate exemplary embodiments of the invention and is not intended to represent the only embodiments in which the invention may be practiced.

These embodiments are provided solely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the present invention of the scope of the invention, and that the present invention will be defined by the scope of the claims. It's just that.

In some cases, in order to avoid ambiguity of the concept of the present invention, well-known structures and devices may be omitted or may be shown in block diagram form focusing on the core functions of each structure and device. In addition, the same components are described using the same reference numerals throughout this specification.

Throughout the specification, when a part is said to “comprise or include” a certain element, this means that it does not exclude other elements but may further include other elements, unless specifically stated to the contrary. do.

Additionally, the term "unit" used in the specification refers to a unit that processes at least one function or operation, and may be implemented as hardware, software, or a combination of hardware and software. Furthermore, the terms “a” or “an”, “one”, and similar related terms may be used in the singular and plural in the context of describing the present invention, unless otherwise indicated herein or clearly contradicted by context. It can be used in a meaning that includes.

In addition, specific terms used in the embodiments of the present invention are provided to aid understanding of the present invention, and unless otherwise defined, all terms used herein, including technical or scientific terms, refer to the present invention. It has the same meaning as generally understood by those with ordinary knowledge in the technical field to which it belongs. The use of these specific terms may be changed to other forms without departing from the technical spirit of the present invention.

Hereinafter, embodiments of a method, device, or system for automatically diagnosing and predicting abnormal functions and vulnerabilities in a bootloader through a deep-learning algorithm according to the present invention will be described. Explain.

Hereinafter, in this specification, to help understand the present invention and for convenience of explanation, the firmware bootloader is described as an example, but the present invention is not limited thereto, and includes a firmware device driver, a device It can also be applied to small operating systems (OS) and application software running on the device. That is, the present invention can be applied to a system that diagnoses firmware.

Various embodiments of the present invention will be described with reference to the attached drawings.

Figure 2 shows a schematic configuration of an on-device AI equipment artificial intelligence engine security abnormal behavior detection and testing method according to an embodiment of the present invention.

Referring to Figure 2, 1. On-device AI equipment has an SoC that supports an artificial intelligence engine, can directly control the SoC's artificial intelligence operation program, and operates a program that provides electromagnetic wave side channel information for learning. May contain collectible equipment.

Additionally, 2. On-device AI equipment using the same SoC is an actual test subject using the SoC that supports the artificial intelligence engine used for learning, and may include equipment that cannot directly operate the SoC's artificial intelligence operation program. .

Meanwhile, 3. SoC's artificial intelligence engine operation program may include a program that can directly control the SoC's artificial intelligence operation program, and is system software that can directly control the hardware of the artificial intelligence engine, and is capable of controlling the artificial intelligence engine operation. It can be used to measure electromagnetic wave side channel characteristics.

4. The artificial intelligence operation program shown in Figure 2 is a program that is mounted on on-device AI equipment and performs artificial intelligence functions, and is installed or sold in program form by equipment manufacturers that do not provide source code and detailed information. program that works. Basically, learning and classification tasks are performed using the SoC's artificial intelligence engine.

In addition, 5. Sub-channel characteristic measurement measures the electromagnetic wave side channel generated from the SoC or power supply of the on-device AI equipment, converts it into a file in the I/Q data format of the radio wave characteristics using a spectrum analyzer, and generates Data can be converted into picture format to perform deep learning learning and classification.

6. Abnormal security behavior in Figure 2 monitors on-device AI equipment communication, equipment log information collection, external access control, etc. to monitor security such as abnormal communication, abnormal behavior through log information, abnormal login, and external login attempts. Abnormal behavior patterns can be monitored through real-time monitoring and history inquiry.

In addition, 7. Characteristic deep learning learning can perform deep learning learning on the operational data set for each function of the artificial intelligence engine created through side-channel characteristic measurement while operating the artificial intelligence engine on the SoC of the on-device AI equipment.

8. Characteristic deep learning classification in Figure 2 can classify each operation of the artificial intelligence engine when operated on on-device AI equipment using the same SoC, based on the learned characteristics of the artificial intelligence engine.

In addition, 9. Association analysis examines the relationship between the operational characteristics of the artificial intelligence engine classified by deep learning and security abnormal behavior, and examines the relevance of security abnormal behavior related to the operation of the artificial intelligence engine to determine the security of the relevant on-device AI. Vulnerabilities can be tested.

10. On-device AI security abnormal behavior detection and testing system in Figure 2 is a system that configures a system that performs each of the operations 5 to 9 above and serves as security verification test equipment and detection of security abnormal behavior for on-device AI. can be used

The present invention proposes an artificial intelligence engine operation program of 3. SoC that can measure the electromagnetic wave side channel characteristics of on-device AI equipment using a SoC including an artificial intelligence engine, and configures the characteristics of the artificial intelligence engine as a dataset. When conducting learning and performing artificial intelligence operations based on the learned artificial intelligence engine characteristics for various types of on-device AI equipment using the same SoC, it is necessary to determine whether the artificial intelligence engine is used and security abnormalities of the equipment. The goal is to provide a system that can detect security abnormalities in artificial intelligence and perform security tests on artificial intelligence equipment by performing correlation analysis of behavior.

Figure 3 is a flow chart configuring the electromagnetic side channel characteristic learning data set for the artificial intelligence engine in the standard equipment, and Figure 4 shows the same artificial intelligence engine based on the operation of the artificial intelligence engine learned based on deep learning. This is a flowchart that detects security abnormalities in the artificial intelligence operation of various types of on-device AI equipment used in the test room and field while operating them.

Table 1 and Table 2 below show the content and flow of each configuration used in 3. SoC's artificial intelligence engine operation program. In this program, while performing operations for each case, electromagnetic wave side channel information for the artificial intelligence engine can be collected to form a dataset used for learning.

general exam Test Items Detail operate Artificial intelligence engine opened
Artificial intelligence engine close Encoder/Decoder Encoder processing
Decoder processing event Basic event function processing
- signal
- atmosphere
- reset
- Check status cue Queue function function
- Cue -> Insert
- Cue -> Subtract
- Check queue status semaphore Semaphore wait
Semaphore state
Add semaphore
Remove semaphore tread test Trad start
Trad End
Check tread condition time test Check current time
sleep function

functional test test Detail Asynchronous processing function Asynchronous processing function
0: variable definition
1: Initialize artificial intelligence engine
2: Preparing the artificial intelligence engine
3: Artificial intelligence engine asynchronous processing
4: Artificial intelligence engine stops
5: Removal of artificial intelligence engine
6: Removal of related functions Artificial intelligence engine operation removal Actions related to removing artificial intelligence engine actions
0: variable definition
1: Initialize artificial intelligence engine
2: Removal of artificial intelligence engine initialization test 0: variable definition
1: Initialize artificial intelligence engine Plugin Management Plug-in operations related to the operation of the artificial intelligence engine
- Plugin management test
1: Start the Plugin Manager
2: Quit Plugin Manager
3: Remove plugin staging function Skills tests associated with preparation
0: variable definition
1: Initialize artificial intelligence engine
2: Preparing the artificial intelligence engine Artificial intelligence engine stops Artificial intelligence engine stop function test
0: variable definition
1: Initialize artificial intelligence engine
2: Preparing the artificial intelligence engine
3: Artificial intelligence engine stops Select setting function Select (Settings/Status Check) Setting Function
0: variable definition
1: Initialize artificial intelligence engine
2: Preparing the artificial intelligence engine
3: Select function settings
4. Check selected functions Synchronous processing function Synchronous processing function test
0: variable definition
1: Initialize artificial intelligence engine
2: Preparing the artificial intelligence engine
3: Artificial intelligence engine synchronous processing
4: Artificial intelligence engine stops
5: Removal of artificial intelligence engine
6: Removal of related functions

When performing each functional operation related to the operation of the artificial intelligence engine shown in Table 1 and Table 2, the electromagnetic wave data generated from the SoC is classified by each item to form a dataset.

The purpose of this process is to perform deep learning on the constructed dataset to learn the operation details of each state of the artificial intelligence engine.

Using the same SoC, electromagnetic wave side channel characteristics are collected for the operation of the internal artificial intelligence engine for equipment that is unknown to the tester or in the actual field as to how the program is configured, and a dataset is constructed in real time. When deep learning is configured to perform classification, the operating status of the artificial intelligence engine of the target device is checked, and the hardware malware function of programs and devices that use the artificial intelligence engine is checked by performing correlation analysis with tools that detect abnormal security behavior. and provide a system to check security vulnerabilities and detect unknown external attacks.

The pattern learning and abnormal behavior classification method for electromagnetic wave emissions of the on-device AI SoC neural network acceleration engine of the present invention can be used in various pattern learning and abnormal behavior classification methods.

Claims (1)

Measuring the electromagnetic wave side channel generated from the power supply of the on-device AI equipment;
Generating the measured electromagnetic wave side channel as data in a radio wave characteristic data format file;
Converting the generated data into picture format to perform deep learning learning and classification; and
Including the step of examining the correlation between the operating characteristics of the artificial intelligence engine classified by deep learning and abnormal security behavior,
The inspection step is characterized by testing security vulnerabilities of the on-device AI by examining the relevance of abnormal security behavior associated with the operation of the artificial intelligence engine. Pattern learning for electromagnetic wave emissions from the on-device AI SoC neural network acceleration engine. and abnormal behavior classification methods.