CN111416735B - Federal learning-based safety QoS prediction method under mobile edge environment - Google Patents

Federal learning-based safety QoS prediction method under mobile edge environment Download PDF

Info

Publication number
CN111416735B
CN111416735B CN202010135672.9A CN202010135672A CN111416735B CN 111416735 B CN111416735 B CN 111416735B CN 202010135672 A CN202010135672 A CN 202010135672A CN 111416735 B CN111416735 B CN 111416735B
Authority
CN
China
Prior art keywords
edge
public
training
qos
data set
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010135672.9A
Other languages
Chinese (zh)
Other versions
CN111416735A (en
Inventor
金惠颖
张鹏程
吉顺慧
李清秋
张雅玲
魏芯淼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN202010135672.9A priority Critical patent/CN111416735B/en
Publication of CN111416735A publication Critical patent/CN111416735A/en
Application granted granted Critical
Publication of CN111416735B publication Critical patent/CN111416735B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/147—Network analysis or design for predicting network behaviour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/28—Restricting access to network management systems or functions, e.g. using authorisation function to access network configuration
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/50—Network service management, e.g. ensuring proper service fulfilment according to agreements

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Security & Cryptography (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Telephonic Communication Services (AREA)

Abstract

The invention discloses a security QoS prediction method under a mobile edge environment based on federal learning. The method comprises the steps of firstly, collecting edge position information and a QoS data set, and fusing to obtain a time-space edge user QoS data set; then, the edge area is divided, a public data set is extracted, and public model training and personalized prediction are combined to achieve the purpose of safety. In the public model training process, the geographical position distribution of the edge server and the municipal administration areas is considered, the whole edge network area is divided into a plurality of edge areas, and each edge area corresponds to a plurality of municipal administration areas. And after public data extraction is carried out on each edge area, public model training is carried out based on the LSTM, and public weight parameters are transmitted to the private users. In the personalized prediction process, the user uses the public LSTM weight parameter in the area to which the user belongs as the initial parameter of the private LSTM, and the private LSTM is continuously trained according to the private data of the user so as to perform personalized prediction.

Description

Federal learning-based safety QoS prediction method under mobile edge environment
Technical Field
The invention relates to a QoS prediction method, in particular to a safe QoS prediction method under a mobile edge environment, and belongs to the technical field of information.
Background
SOA (service Oriented architecture) is an application architecture in which all functions are defined as independent services, and Web services are one of the technologies for implementing the SOA, so that industry experts provide users with ever-changing demands by combining the services. In recent years, with the development of Web services, the non-functional attribute qos (quality of service) of the Web services has received more and more attention, and nowadays, a large number of Web services with the same or similar functions appear on the network, so it is important to select an appropriate Web service to meet the needs of users. On the other hand, with the advent of the 5G era, mobile edge computing has become more widely used, and since it is located at the edge of the network, in close proximity to users or information sources, the latency of responding to requests can be greatly reduced. Providing service response for users in the mobile edge environment is a current trend, but also brings security problems in the mobile edge environment.
The existing QoS security research work focuses mainly on two aspects: QoS privacy protection method and QoS security service. In terms of QoS privacy protection, the earliest scholars proposed a differential privacy encryption algorithm, and subsequently liu et al combined it with a collaborative filtering method to propose a QoS privacy protection method. Qi et al in 2017 proposed a locality-sensitive-hash-based distributed recommendation system. Shahriar et al propose a protection protocol for attribute value encryption and location hiding. Therefore, the existing QoS privacy protection method is mostly suitable for static environment, and the prediction precision is damaged after data encryption. On the other hand, since users predict by accessing each other's history data in the marginal environment, the data encryption rule is more easily broken as the number of interactions increases.
In terms of QoS security services, researchers have proposed a variety of security mechanisms and criteria. Shen et al, which considers both security and quality of service, propose a distributed dynamic management system mechanism, but the mechanism is only applicable in specific environments (e.g., when network traffic is lightweight). Alessandro et al propose an integrated tool support method that achieves the greatest tradeoff between security and quality, but that does not take into account the changing conditions in a dynamic environment. Jalal et al use a method of security QoS optimization in a distributed real-time environment to agree on confidentiality, integrity and authentication security, but this method is mostly used for IP routing protocols. Charuenporn et al propose a new QoS security metric development paradigm, however, this paradigm is applicable only to two defined information system standards (COBIT and ITIL).
With the development of technology, a great deal of private information of users is acquired in more and more prediction methods, and safety becomes a great demand for users in the prediction process.
Disclosure of Invention
The purpose of the invention is as follows: considering that the traditional method is not applicable under the mobile edge environment, the prediction precision is greatly reduced and the requirement of a user on safety is met, the invention provides a Federal learning-based safety QoS prediction method under the mobile edge environment. Meanwhile, the LSTM model is used for dynamically updating the weight parameters to improve the prediction precision, so that the purpose of effective and accurate prediction is achieved while the safety is ensured.
The technical scheme is as follows: in order to achieve the above object, the method for predicting the security QoS in the mobile edge environment based on federal learning according to the present invention includes the following steps:
step 1: collecting edge position information and a QoS data set;
step 2: fusing QoS data and edge position points by taking the user ID number as connection;
and step 3: obtaining a time-space edge user QoS data set after fusion;
and 4, step 4: the geographical position distribution of an edge server and a city district is considered, an original edge network area is divided into a plurality of edge areas, and public data set extraction is carried out;
and 5: performing public data training based on LSTM by utilizing the public data sets of all the edge areas to obtain a public model;
step 6: and taking the public LSTM weight parameter in the area to which the user belongs as an initial parameter of the private LSTM, and continuously training the private LSTM according to the private data so as to perform personalized prediction.
Preferably, the data collected in step 1 mainly includes two aspects: a QoS data set containing a user ID, a service ID, a time period ID, and an attribute value; a base station data set containing latitude and longitude information.
Preferably, the step 2 comprises the following steps:
step 21: sorting the QoS data sets in the order of the user ID, the time period ID, the service ID and the attribute value;
step 22: randomly selecting a corresponding number of edge position points in the base station data set according to the total number of the users in the step 21 and carrying out ID numbering;
step 23: counting longitude and latitude information of the edge position points, and positioning the position points of the base station through map service;
step 24: the total number of users in the QoS dataset and the total number of edge servers are equal, so the two datasets are merged with the user ID as the connection.
Preferably, the step 3 comprises the following steps:
step 31: counting the total time segment number of the fused data set and the total number of formed edge servers;
step 32: the total time period is abbreviated as 'hour', the total edge server is abbreviated as 'null', and a time-null edge user QoS data set is obtained after fusion.
Preferably, the step 4 comprises the following steps:
step 41: considering the geographical position distribution of an edge server and a municipal district, dividing the whole edge network area into k edge areas, wherein each edge area corresponds to 1-4 municipal districts, and k is more than or equal to 2 and is determined by the geographical distribution of the edge server;
step 42: taking out each edge region T1All users in the time period call attribute values of all services, and the attribute values are expressed in a service-user two-dimensional matrix form; the median of each row of data in the two-dimensional matrix is taken one by one to obtain a service-T1A column vector, which is taken as T1Public data of attribute values of the service called by the time period;
step 43: take T in sequence2、…、TnAll users in the time period call the attribute values of all services, and step 42 is repeated to obtain service-T2…, service-TnColumn vectors, where n ≧ 2;
step 44: and synthesizing the n column vectors to form a service-time period two-dimensional matrix for the public model training of the edge area.
Preferably, the step 5 comprises the following steps:
step 51: each edge region performs LSTM training based on the respective region's common data set; a consensus is reached between learning rate and training times, namely: the larger the learning rate is, the faster the error adjustment speed is, so the training times required to be executed when the same training effect is achieved are less;
step 52: setting the learning rate and the training times as a group of combination parameters for public model training, wherein the training times are gradually reduced when the learning rate is increased due to the negative correlation relationship between the learning rate and the training times;
step 53: calculating the final output through a forgetting gate, an input gate, the current time unit state and an output gate in the LSTM, and calculating a root mean square error value to measure a loss function;
step 54: deriving the loss function and continuously adjusting the weight parameter; with the increase of the training iteration times, the training error is continuously reduced, and the optimal initial weight parameter is provided for personalized prediction.
Preferably, in the personalized prediction process, the step 6 executes LSTM training at intervals based on private data continuously generated by the user, so as to continuously update the weight parameter to meet the requirements of the edge environment on the real-time performance and accuracy of the data, thereby realizing the prediction of the safe QoS attribute value in the edge environment.
Preferably, the step 6 comprises the following steps:
step 61: the user carries out QoS prediction according to the initial weight parameter transmitted in the public LSTM;
step 62: after the prediction of a time period is completed, adjusting model weight parameters according to the deviation of the predicted data and the actual QoS data;
and step 63: judging whether new data is generated in the next time period, if so, transmitting the latest weight parameter to the model prediction of the time period, and executing training so as to further adjust the weight parameter; therefore, the steps are repeated until new data are not generated any more, and the real-time personalized prediction is realized.
Has the advantages that: compared with the prior art, the method for predicting the safety QoS under the mobile edge environment based on the federal learning provided by the invention has the advantages that on one hand, the defect that the traditional encryption mode is easy to crack under the edge environment is overcome by combining the public model training and the personalized prediction, and on the other hand, the prediction precision under the edge environment is ensured by continuously updating the weight parameters while the safety is ensured.
Drawings
FIG. 1 is an overall step diagram of an embodiment of the present invention;
FIG. 2 is a flow chart of a method according to an embodiment of the present invention.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
As shown in fig. 1, the method for predicting the security QoS in the mobile edge environment based on federal learning disclosed in the embodiment of the present invention mainly includes 6 steps:
step 1: collecting an edge position and QoS data set with longitude and latitude information;
step 2: corresponding the edge position point and the QoS data set one by one through a user ID number and fusing the edge position point and the QoS data set;
and step 3: obtaining a QoS data set with geographic position attributes after fusion, and obtaining a time-space edge user QoS data set based on the mapped edge position points;
and 4, step 4: the geographical position distribution of an edge server and a municipal district is considered, an original edge network area is divided into a plurality of edge areas, and a common data set of each area is extracted for common model training;
and 5: carrying out public data training on the basis of the LSTM by utilizing the public data sets of all the areas to obtain a public model, and transmitting a public weight parameter to a private user;
step 6: and taking the public LSTM weight parameter in the area to which the user belongs as an initial parameter of the private LSTM, and executing private LSTM training based on private data so as to continuously update the weight parameter for personalized prediction.
As shown in fig. 2, the method for predicting the security QoS in the mobile edge environment based on federal learning disclosed in the embodiment of the present invention includes the following specific steps:
step 101: and collecting edge position information and a QoS data set, wherein the edge position information mainly comprises the latitude and longitude of the base station, and the QoS data set mainly comprises representative attributes such as response time, throughput and the like.
In this step, a QoS data set including a user ID, a service ID, a time period ID, and an attribute value is collected from wsdream, which is a channel for collecting user information, and can provide Web service reliability and quality evaluation. Counting the total time period number of the QoS data set with the time attribute according to the time period ID number, wherein the user IDs in the data set are used for corresponding to the IDs of the edge position points one by one; a shanghai telecom base station data set containing latitude and longitude information is collected from the telecom, which contains 720 million internet access records of 9481 mobile phones collected by 3233 base stations.
Step 102: rearranging the QoS data set according to the order of the user ID, the time period ID, the service ID and the attribute value; the data size of the QoS data set in this step is 142 × 4500 × 64, that is, the attribute value data set containing response time and throughput formed by 142 users invoking 4500 services in 64 time periods. The QoS data sets are rearranged in the order of the user ID (0-141), the period ID (0-63), the service ID (0-4499), and the attribute value.
Step 103: randomly selecting a corresponding number (142 groups) of edge position points in the data set of the Shanghai telecommunication base station according to the total number (142) of users in the step 102 and carrying out ID numbering (0-141).
Step 104: counting longitude and latitude information of the edge position points, and positioning the position points of the base station through a map service (carefree map) to obtain the distribution condition of the edge server; this example 142 sets of data covers 8 municipalities, Min, Huangpu, irid, Pudong New, Xuhui, Songjiang, Qingpu and Silian, Shanghai.
Step 105: the total number of users in the QoS data set and the total number of edge servers are equal (both 142), so that the two data sets are fused by using user IDs (0-141) as connections and corresponding to the same IDs.
Compared to a conventional QoS dataset, a QoS dataset in an edge environment has spatial attributes, and in addition the dataset has temporal attributes. After the fused data set is mapped, obtaining an edge position point to obtain a time-space edge user QoS data set; the specific operation steps for acquiring the time-space edge user QoS data set are as follows:
step 106: and counting the total number of time segments (64 time segments, the interval of each time segment is 15 minutes) of the fused data set and the total number of formed edge servers (142), wherein the time segment ID (0-63) is abbreviated as 'hour', and counting the number of edge servers formed by positioning 142 groups of longitude and latitude information to the geographic position points, wherein the total number is 67. The geographical location distribution of the edge server is referred to as 'null', so that the two are correspondingly fused to obtain a time-null edge user QoS data set.
The dividing of the edge area and the extraction of the public data set are used for further enhancing the security, and the specific processing steps for dividing the edge area are as follows:
step 107: and considering the geographical position distribution of the edge server and the urban area, dividing the whole edge network area into k (wherein k is more than or equal to 2 and is determined by the geographical distribution of the edge server) edge areas, wherein each edge area corresponds to 1-4 urban areas. In the formed time-space edge user QoS data set, according to the formed 67 edge servers and the distribution situation of the Shanghai urban districts, the urban districts which are closer in geographic position are divided into an edge area because the edge servers which are closer in geographic position have a more similar edge environment. The Shanghai area is divided into three edge areas, edge area 1: minkout, 22 edge servers, edge area 2: huangpu zone, iris zone and purdong new zone, 24 edge servers, edge zone 3: xu hui district, Songjiang district, Qingpu district and quiet district, 21 edge server.
Step 108: taking out each region T1All users invoke attribute values of all services in a time period, wherein the attribute values comprise response time and throughput, and service-user twoA formal representation of the dimensional matrix. Such as: taking attribute values of all services called by all users in the time period 1 of the edge area 1 to form a T1A service-user matrix of a time period, and the median of each row of data in the two-dimensional matrix is taken one by one to obtain a service-T1A column vector, which is taken as T1The user of the time slot edge area 1 calls public data of the attribute value of the service; then taking the T of the edge region 1 in sequence2、…、T64All users in the time period call attribute values of all services to obtain service-T of the edge area 12service-T64A column vector. The edge region 2 and the edge region 3 can obtain service-time period column vectors in the same way; the 64 column vectors of the edge area 1 are synthesized to form a service-time period two-dimensional matrix, and the edge area 1 forms two common data sets of response time-time period and throughput-time period for common model training of the area. Edge region 2 and edge region 3 can likewise form a common data set by synthesis.
The public model training based on the LSTM can provide initial weight parameters for private users, and the specific processing steps of the public data training are as follows:
step 109: each edge region performs LSTM training based on the public dataset of the two-dimensional matrix of response time (rt) and throughput (tp) of the respective region. A consensus is reached between learning rate and training times, namely: the larger the learning rate is, the faster the error adjustment speed is, so the training times required to be executed when the same training effect is achieved are less; the learning rate and the training times are set as a group of combination parameters for public model training, and due to the negative correlation relationship between the learning rate and the training times, the value of the training times is gradually reduced when the value of the learning rate is increased. E.g. learning rate in the interval 0.001,0.01 in steps 0.001]Taking the value above, the corresponding training times are in the interval [100,1000 ] by the step length of 100]The upper and lower values are decreased, so that the balance between the learning rate and the training times is achieved; after parameter setting is completed, final output is calculated through a forgetting gate, an input gate, a unit state at the current moment and an output gate in the LSTM, and a loss function is estimated through calculation of a root mean square error value. LSTM inputs QoS attribute values for each service at a previous time using a previous time(response time or throughput), predicting the QoS attribute value of each service at the next moment, wherein the specific model is represented as: (1) f. oft＝σ(Wf·[ht-1,xt]+bf) Wherein W isfIs the weight matrix of the forgetting gate, ht-1And xtRespectively representing the output of the previous time and the input of the current time, bfIs a biased term for a forgetting gate. The LSTM model selectively reserves the unit state from the previous moment to the unit state from the current moment through a forgetting gate; (2) i.e. it＝σ(Wi·[ht-1,xt]+bi) Wherein W isiIs a weight matrix of the input gate, biIs an offset item of an input gate that selectively saves the input cell state at the current time; (3) calculating the cell state c at the current timet. It is determined from the cell state c at the previous timet-1Multiplication by element of forget gate ftReuse the currently input cell state
Figure BDA0002397224430000071
Multiplying input Gate i by elementtFinally, the two products are added
Figure BDA0002397224430000072
(4) By (3) memorizing the LSTM with respect to the current
Figure BDA0002397224430000073
And long term memory ct-1Combine to form a new cell state ct. Due to the control of the forgetting gate and the input gate, they can save information of a long time ago and avoid the current irrelevant contents from entering the memory. ot＝σ(Wo·[ht-1,xt]+bo) Is a calculation process of an output gate, which controls the influence of long-term memory on the current output; (5) the final output of the LSTM is determined by the output gate and the cell state together ht＝otοtanh(ct) (ii) a (6) Calculating a root mean square error value
Figure BDA0002397224430000074
To measure the loss functionA number of, wherein
Figure BDA0002397224430000075
Is tiThe output value at the time of the instant LSTM,
Figure BDA0002397224430000076
is tiThe true value of the time, N is the total number of times.
Step 110: and obtaining a loss function of each prediction result, and continuously adjusting the weight parameter by differentiating the loss function. Therefore, with the increase of training iteration times, training errors are continuously reduced, weight parameters are gradually accurate, and finally 2 common models trained by the response time data set and the throughput data set in each region are obtained.
Step 111: and after the public model training is finished in each edge area, the public weight parameter is stored, and the weights of a forgetting gate, an input gate, a unit state and an output gate are provided for the private users in the area as the optimal initial weight parameter for personalized prediction.
Step 112: the user in each region performs QoS prediction according to the initial weight parameters passed in the region public LSTM, and the calculation process is as described in (1) - (5) of step 109, and T1,T2,…,T64And predicting the sequence of the time periods and obtaining a prediction result. And the weight parameters in the prediction model are continuously updated based on the derivation of the prediction error.
Step 113: when T is completed1After the time period is predicted, the weight parameter of the model is adjusted according to the deviation of the predicted data and the actual QoS data, and the weight parameter adjusting formula is as follows: weighti＝weighti-l error, where weightiIs the initial weight parameter, l is the learning rate,
Figure BDA0002397224430000081
is to take the derivative of the prediction error,
Figure BDA0002397224430000082
is tiThe output value at the time of the instant LSTM,
Figure BDA0002397224430000083
is tiThe true value of the time of day.
Step 114: the model is updated once each time training is completed for a time interval (which can be set according to system requirements and data conditions, such as 1 hour, 1 day, etc. in specific applications), the model at this moment is called a private model, and weight parameters in the model are only private for the user. Along with the continuous iteration of the training times, the private model is continuously perfected, and the weight parameters in the model are continuously updated.
Step 115: and when the training is finished each time, judging whether new data is generated in the next time period, if so, continuing to perform private prediction according to the method in the step 112, and obtaining a prediction result in the time period. If not, the training is finished.
Step 116: and outputting a final prediction result after the prediction and training of all time periods are finished.

Claims (5)

1. A method for predicting the safety QoS under the mobile edge environment based on the federal learning is characterized by comprising the following steps:
step 1: collecting edge position information and a QoS data set;
step 2: fusing QoS data and edge position points by taking the user ID number as connection;
and step 3: obtaining a time-space edge user QoS data set after fusion;
and 4, step 4: the geographical position distribution of an edge server and a city district is considered, an original edge network area is divided into a plurality of edge areas, and public data set extraction is carried out;
and 5: performing public data training based on LSTM by utilizing the public data sets of all the edge areas to obtain a public model;
step 6: taking public LSTM weight parameters in the area to which the user belongs as initial parameters of the private LSTM, and continuously training the private LSTM according to private data so as to perform personalized prediction;
the data collection in the step 1 mainly comprises two aspects: a QoS data set containing a user ID, a service ID, a time period ID, and an attribute value; a base station data set containing latitude and longitude information;
the step 2 comprises the following steps:
step 21: sorting the QoS data sets in the order of the user ID, the time period ID, the service ID and the attribute value;
step 22: randomly selecting a corresponding number of edge position points in the base station data set according to the total number of the users in the step 21 and carrying out ID numbering;
step 23: counting longitude and latitude information of the edge position points, and positioning the position points of the base station through a map service to obtain the distribution condition of the edge server;
step 24: the total number of users in the QoS data set is equal to the total number of edge servers, so that the two data sets are fused by taking the user ID as connection;
the step 4 comprises the following steps:
step 41: considering the geographical position distribution of an edge server and a municipal district, dividing the whole edge network area into k edge areas, wherein each edge area corresponds to 1-4 municipal districts, and k is more than or equal to 2 and is determined by the geographical distribution of the edge server;
step 42: taking out each edge region T1All users in the time period call attribute values of all services, and the attribute values are expressed in a service-user two-dimensional matrix form; the median of each row of data in the two-dimensional matrix is taken one by one to obtain a service-T1A column vector, which is taken as T1Public data of attribute values of the service called by the time period;
step 43: take T in sequence2、…、TnAll users in the time period call the attribute values of all services, and step 42 is repeated to obtain service-T2…, service-TnColumn vectors, where n ≧ 2;
step 44: and synthesizing the n column vectors to form a service-time period two-dimensional matrix for the public model training of the edge area.
2. The method of claim 1, wherein the step 3 comprises the following steps:
step 31: counting the total time segment number of the fused data set and the total number of formed edge servers;
step 32: the total time period is abbreviated as 'hour', the total edge server is abbreviated as 'null', and a time-null edge user QoS data set is obtained after fusion.
3. The method of claim 1, wherein the step 5 comprises the following steps:
step 51: each edge region performs LSTM training based on the respective region's common data set; a consensus is reached between learning rate and training times, namely: the larger the learning rate is, the faster the error adjustment speed is, so the training times required to be executed when the same training effect is achieved are less;
step 52: setting the learning rate and the training times as a group of combination parameters for public model training, wherein the training times are gradually reduced when the learning rate is increased due to the negative correlation relationship between the learning rate and the training times;
step 53: calculating the final output through a forgetting gate, an input gate, the current time unit state and an output gate in the LSTM, and calculating a root mean square error value to measure a loss function;
step 54: deriving the loss function and continuously adjusting the weight parameter; with the increase of the training iteration times, the training error is continuously reduced, and the optimal initial weight parameter is provided for personalized prediction.
4. The method of claim 1, wherein in the step 6, in the personalized prediction process, the LSTM training is performed every other time interval based on the private data continuously generated by the user, so as to continuously update the weight parameters to meet the requirements of the edge environment on the real-time performance and accuracy of the data, thereby realizing the prediction of the QoS attribute values in the edge environment.
5. The method of claim 4, wherein the step 6 comprises the following steps:
step 61: the user carries out QoS prediction according to the initial weight parameter transmitted in the public LSTM;
step 62: after the prediction of a time period is completed, adjusting model weight parameters according to the deviation of the predicted data and the actual QoS data;
and step 63: judging whether new data is generated in the next time period, if so, transmitting the latest weight parameter to the model prediction of the time period, and executing training so as to further adjust the weight parameter; therefore, the steps are repeated until new data are not generated any more, and the real-time personalized prediction is realized.
CN202010135672.9A 2020-03-02 2020-03-02 Federal learning-based safety QoS prediction method under mobile edge environment Active CN111416735B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010135672.9A CN111416735B (en) 2020-03-02 2020-03-02 Federal learning-based safety QoS prediction method under mobile edge environment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010135672.9A CN111416735B (en) 2020-03-02 2020-03-02 Federal learning-based safety QoS prediction method under mobile edge environment

Publications (2)

Publication Number Publication Date
CN111416735A CN111416735A (en) 2020-07-14
CN111416735B true CN111416735B (en) 2021-05-11

Family

ID=71492750

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010135672.9A Active CN111416735B (en) 2020-03-02 2020-03-02 Federal learning-based safety QoS prediction method under mobile edge environment

Country Status (1)

Country Link
CN (1) CN111416735B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111836321B (en) * 2020-07-27 2022-11-18 北京邮电大学 Cell switching method based on federal learning and edge calculation
CN113099387B (en) * 2021-03-22 2022-02-22 河海大学 Method for constructing service quality data set in real edge environment
CN113194489B (en) * 2021-04-01 2022-09-13 西安电子科技大学 Minimum-maximum cost optimization method for effective federal learning in wireless edge network
CN113420421B (en) * 2021-05-28 2023-07-25 西安邮电大学 QoS prediction method based on time sequence regularized tensor decomposition in mobile edge calculation

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108900346A (en) * 2018-07-06 2018-11-27 西安电子科技大学 Wireless network traffic prediction technique based on LSTM network
CN109981333A (en) * 2018-12-28 2019-07-05 华为技术有限公司 A kind of O&M method and O&M equipment applied to data center
CN110191015A (en) * 2019-05-21 2019-08-30 中山大学 Cloud service performance intelligent Forecasting and device based on CPI index
CN110458336A (en) * 2019-07-23 2019-11-15 内蒙古工业大学 A kind of net based on deep learning about vehicle supply and demand prediction method

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3804226A1 (en) * 2018-06-06 2021-04-14 The Joan and Irwin Jacobs Technion-Cornell Institute Telecommunications network traffic metrics evaluation and prediction
CN110247793B (en) * 2019-05-29 2022-11-11 暨南大学 Application program deployment method in mobile edge cloud

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108900346A (en) * 2018-07-06 2018-11-27 西安电子科技大学 Wireless network traffic prediction technique based on LSTM network
CN109981333A (en) * 2018-12-28 2019-07-05 华为技术有限公司 A kind of O&M method and O&M equipment applied to data center
CN110191015A (en) * 2019-05-21 2019-08-30 中山大学 Cloud service performance intelligent Forecasting and device based on CPI index
CN110458336A (en) * 2019-07-23 2019-11-15 内蒙古工业大学 A kind of net based on deep learning about vehicle supply and demand prediction method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于用户行为和位置感知的边际服务加载优化研究;童智高;《中国优秀硕士学位论文全文数据库》;20171115(第11期);全文 *

Also Published As

Publication number Publication date
CN111416735A (en) 2020-07-14

Similar Documents

Publication Publication Date Title
CN111416735B (en) Federal learning-based safety QoS prediction method under mobile edge environment
Liu et al. Dynamic spatial-temporal representation learning for traffic flow prediction
Guo et al. Deep federated learning enhanced secure POI microservices for cyber-physical systems
Xu et al. A survey for mobility big data analytics for geolocation prediction
Kong et al. Construction of intelligent traffic information recommendation system based on long short-term memory
WO2017185664A1 (en) Method for positioning terminal, and network device
Zhang et al. Quality-aware user recruitment based on federated learning in mobile crowd sensing
Yu et al. Modeling user activity patterns for next-place prediction
Chen et al. An approach of flow compensation incentive based on Q-learning strategy for IoT user privacy protection
CN111309780B (en) Track data processing method and device
WO2022198947A1 (en) Method and apparatus for identifying close-contact group, and electronic device and storage medium
EP3425876A1 (en) Location-based service implementing method and apparatus
CN111401743A (en) Dynamic traffic influence evaluation method in urban road construction period
EP3192061B1 (en) Measuring and diagnosing noise in urban environment
Yuan et al. Deep multi-view residual attention network for crowd flows prediction
WO2020223908A1 (en) Privacy management
CN104123321A (en) Method and device for determining recommended pictures
Ma et al. Cellular traffic prediction via deep state space models with attention mechanism
CN108009134A (en) A kind of Information of Ancient Human Activity method for digging
CN112287503A (en) Dynamic space network construction method for traffic demand prediction
Li et al. A multi-task scheduling mechanism based on ACO for maximizing workers’ benefits in mobile crowdsensing service markets with the Internet of Things
CN109581280A (en) The adaptive tuning on-line method, system and device of terminal
US11050834B1 (en) Method for automatically assigning visits to partially observable location data streams
Tang et al. Region-aware POI recommendation with semantic spatial graph
CN103886739A (en) Personalized service platform based on large-scale real-time traffic index system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant