A white traffic filtering model training method and system based on multi-party split learning

Abstract

The present invention discloses a white flow filtering model training method and system based on multi-party split learning, comprising the following steps: step one, data preprocessing; step two, white flow filtering model splitting; step three, white flow filtering model training. The present invention adopts a multi-party collaborative learning strategy, breaks the data island, solves the problem of less labeled data, greatly expands the data set, and can effectively improve the accuracy and robustness of the white flow filtering model; adopts a computing power level evaluation strategy, which can effectively solve the problem of large differences in computing power among many training parties. Dividing the model after evaluation can ensure that all training parties can successfully complete the training of the model within the specified time. Generally, a hidden layer is allocated to the training party, which greatly saves the computing resources of the training party.

Description

Technical Field

[0001] The present invention relates to the field of traffic filtering, and in particular to a white traffic filtering model training method and system based on multi-party split learning. Background Art

[0002] With the advancement of Internet technology, application-layer traffic on networks is becoming increasingly complex. From minor issues like website slowdowns to major issues like information leaks, identity theft, and commercial losses, malicious traffic has become a growing concern across every industry.

[0003] Currently, traditional security devices use different methods to defend against network traffic. What they have in common is that they are all based on full network analysis. During the detection process, a list is set to match known threats. However, the set list is only valid the next time after the attack, and it cannot quickly adapt to and effectively identify traffic variants. Therefore, in order to effectively ensure network security, it is first necessary to identify and filter out a large amount of white traffic in the network.

[0004] Currently, the most common white traffic filtering methods are: port identification-based methods, deep packet inspection-based classification methods, and statistics-based machine learning methods.

[0005] Port-based identification: In the early days of the internet, random port strategies were not widely used. Researchers simply extracted the port number from the packet header of network traffic and then searched the application table to identify the specific application to which the traffic belonged. However, with the development of the internet, this method is no longer applicable.

[0006] Deep packet inspection-based: This method relies on predefined strings. The entire packet is inspected, and if the predefined string is found anywhere, the traffic type can be determined. While this method is highly accurate, matching the predefined string is time-consuming and cannot handle encrypted traffic.

[0007] Statistical machine learning methods: With the advent of machine learning, computers can perform a wider range of more complex tasks. This method uses statistics to analyze traffic characteristics across various applications, generating statistical information. Models such as decision trees or support vector machines are then selected for training. This method can be used to detect encrypted traffic, but it relies on manual feature design. Due to the growing emphasis on data privacy, these research organizations are limited to using their own data for analysis, leading to the "data silo" problem. Malicious traffic is difficult to detect in network traffic. Relying solely on in-house data for research results in inaccurate and less robust models or systems. This also significantly limits the research of many security vendors on malicious traffic. When filtering white traffic, the lack of labeled data during initial training can lead to malicious traffic being present in the filtered white traffic.

[0008] Glossary:

[0009] Split layer: After the model is split, the last layer of the model allocated to the training party

[0010] Crushed data: the output of the split layer

[0011] Trusted third-party server / trusted third party: used for model partitioning, model allocation, and initialization parameter generation.

[0012] Computing server / server: used for model operation and bears most of the computing resources

[0013] White traffic: non-malicious traffic

[0014] White traffic filtering: Before malicious traffic detection, if all the traffic to be classified is used as input, the classification time and difficulty will be increased. Moreover, when performing real-time malicious traffic detection, most of the captured traffic is white traffic. This white traffic filtering solution can first filter out most of the white traffic through the model before malicious traffic detection, and then perform more in-depth malicious traffic detection on the remaining suspicious traffic (malicious traffic and traffic that cannot be identified).

[0015] Forward propagation: that is, forward propagation. Summary of the Invention

[0016] The technical problem to be solved by the present invention is to provide a white traffic filtering model training method and system based on multi-party split learning in response to the shortcomings of the existing technology.

[0017] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0018] A white traffic filtering model training method based on multi-party split learning includes a trusted third-party server, which is communicatively connected to a computing server and several training parties. The specific steps are as follows:

[0019] Step 1: Data preprocessing: Each training party preprocesses their private data into a unified format;

[0020] Step 2: Split the white traffic filtering model: The trusted third-party server obtains the average computing power level of all training parties, and then divides the white traffic filtering model into split locations based on the average computing power level of all training parties. The location where the white traffic filtering model is split is used as the split layer. The split layer and the previous white traffic filtering model part are distributed to each training party as the pre-training model, and the remaining part is distributed to the computing server as the post-training model.

[0021] Step 3: White traffic filtering model training:

[0022] 3.1) Determine the training order of the training parties;

[0023] 3.2) The first training party inputs its private data into the pre-training model and performs forward propagation to train the pre-training model. The pre-training model outputs fragmented data, which is then transmitted to the computing server. The computing server inputs the fragmented data and corresponding labels into the post-training model and performs forward propagation to train the post-training model to obtain the output results. The computing server calculates the loss function based on the output results and the labels sent by the first training party. The computing server uses the loss function to perform backpropagation on the post-training model and updates the weights using the gradient descent algorithm until it reaches the first layer of the post-training model. The gradient of the first layer of the post-training model is sent to the first training party. The first training party uses the gradient of the first layer of the post-training model to perform backpropagation to update the weights of the first training party's pre-training model to obtain the trained pre-training model. The first training party transmits the model parameters of the trained pre-training model to the next training party.

[0024] 3.3); Each training party repeats step 3.2) until the white flow filtering model converges;

[0025] Step 4: The training party and computing server that converge the white traffic filtering model send the pre-training model parameters and post-training model parameters to the trusted third-party server respectively. The trusted third-party server merges the pre-training model and the post-training model to obtain a complete trained white traffic filtering model and broadcasts it to all training parties.

[0026] Further improvement, in step 1, the training party's private data Where Pdi represents the dataset of training party i, and N represents the number of training parties. Each training party's dataset contains one or more pcap files, and the dataset is divided into files of equal length.

[0027] For further improvement, the SplitCap tool is used to split the pcap file according to the session to generate several small pcap files. The small pcap files are uniformly taken to 784 bytes. For files exceeding 784 bytes, 0x00 is added to the files that are less than 784 bytes. Then the small pcap files are converted into 28×28 grayscale images in png format, and then the grayscale images in png format are converted into IDX file type to become standard input files. Pd i It represents the original data set of training party i, Sd i Indicates the dataset after the original training dataset is converted into images.

[0028] As a further improvement, in step 2, the method for splitting the white traffic filtering model is as follows:

[0029] The trusted third-party server sends a lightweight model and dataset to all trainers. The trainers send the training completion time to the server. The trusted third-party server defines the expected training time: τ, and the training time of trainer i is T i , the computing power coefficient ρ is defined as follows:

[0030]

[0031] When ρ ≥ 1, it means that the average training time of the training party exceeds the expected training time, the computing power is weak, and one hidden layer is allocated to the training party;

[0032] When ρ<1, it means that the average training time of the training party is less than the expected training time, the computing power is strong, and the division hidden layers for the training party, where M represents the number of hidden layers of the model to be trained; Indicates rounding down (1-ρ)M;

[0033] The third-party computing power coefficient divides the white traffic filtering model W, the previous white traffic filtering model W c Sent to the training party, the remaining white traffic filtering model W s The data is sent to the computing server, so that the training party and the computing server cannot know each other's model configuration.

[0034] Further improvement, in step 3, the white traffic filtering model is a deep neural network, the deep neural network is defined as a function F, the deep neural network has an α layer, and the function of running a certain layer is defined as L, then F(Split_data) is equivalent to L α (L α-1 ...(L1(Split_data))), where L α () indicates that forward propagation is performed at layer α; when backpropagation is performed and the gradient is updated, the gradient update operation from layer α to layer 1 is required. The backpropagation of the entire neural network is defined as F T (gradients), the back propagation of the i-th layer is defined as Then F T (gradients) is equivalent to When forward propagation is performed, the parameters of F() are data, gradients represent the gradients of each parameter, and the gradients represent the gradients calculated by the loss function. T represents back propagation; when back propagation is performed, the purpose is to update the gradients, so F T The parameter of () is gradients; the trusted third-party server sends the first β layers to the training party, the β layer is also called the split layer, and the β+1 to α layers are sent to the computing server.

[0035] For further improvement, the white traffic filtering model training steps in step 3 are as follows:

[0036] ① If it is the first training party, it uses the initialization parameters for training. Otherwise, it decrypts the model parameters of the trained model sent by the previous training party for training.

[0037] ②Training party i uses local data for training and propagates W c Train to the last layer, which is the split layer,

[0038] Send the fragmented data and labels to the server, expressed as: L β (L β-1 ...(L1(Sd i )))->F c (Sd i )

[0039] ③ The server receives the fragmented data and labels sent by the trainer, trains Ws to the last layer through forward propagation, and calculates the loss. It can be expressed as:

[0040] L α (L α-1 …(L β+1 (F c (Sd i ))))->F S (F c (Sd i ))

[0041] Loss(F S (F c (Sd i )),label)

[0042] gradients<-Loss′(F S (F c (Sd i )),label)

[0043] Gradients means that after calculating the loss function, the gradients of each parameter are calculated by taking partial derivatives, and then the gradients are updated by gradient descent. S () indicates that the model performs forward propagation on the computing server side, F c Indicates that the model on the training side performs forward propagation. Loss() represents the loss function, that is, the error between the model's running result and the data's own label. Label represents the data label. Loss′() represents the partial derivative of the loss function to calculate the gradient of the parameter.

[0044] ④ The computing server updates W through back propagation s The gradient of each parameter, and W s The gradient of the first layer is sent to the corresponding training party and is expressed as:

[0045]

[0046] in, Indicates that the back propagation is performed on the computing server model. Indicates that back propagation is performed in the β+1 layer, the gradients of the parameters of the β+1 layer are calculated, and then updated;

[0047] ⑤ The training party receives the gradient sent by the computing server and updates the gradient of each parameter through back propagation in Wc, which can be expressed as:

[0048] ⑥ After the training party has updated the gradient, it repeats the above process to complete the training of the local data. After the training is completed, the training party establishes a secure channel with the next training party through a trusted third party and sends the encrypted model parameters of the previous training model to the next training party.

[0049] ⑦ Repeat steps ① to ⑥ until all training methods have completed the training.

[0050] A white flow filtering model training system based on multi-party split learning includes a computer device; the computer device is configured or programmed to execute the steps of the above-mentioned white flow filtering model training method based on multi-party split learning.

[0051] Compared with the existing design methods, the present invention has the following advantages:

[0052] ① This solution uses a multi-party collaborative learning strategy to break down data silos, address the issue of insufficient labeled data, and significantly expand the dataset, effectively improving the accuracy and robustness of the white traffic filtering model. (Before the advent of distributed learning methods, other strategies did not achieve as many data sources as this method.)

[0053] ② This solution utilizes a computing power level assessment strategy, effectively addressing the significant disparity in computing power among multiple training participants. By assigning the model after assessment, it ensures that all participants can successfully complete model training within the specified timeframe. Typically, each participant is assigned one hidden layer, significantly conserving their computing resources.

[0054] ③The key point of this solution is to partition the model, allowing the trainer and server to train separately, so neither party can know the other's model configuration. Even if the server obtains the trainer's model parameters through some attack method, it cannot restore the network structure, thus protecting the trainer's model security and data privacy.

[0055] ④ Research has shown that when the training party uses local data and runs to the split layer, the training party's data cannot be restored through the output of the split layer, which ensures the privacy of the training party in terms of the model.

[0056] ⑤ The trainers use relay training, meaning that a trainer can only begin training after the previous trainer has completed training. Model parameters are also obtained from the previous trainer. From a dataset perspective, this relay training method is equivalent to taking the union of all the datasets of all trainers. It can be thought of as a single trainer training on all white traffic and malicious traffic. This improves data utilization and model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a system architecture diagram of the present invention;

[0058] Figure 2 This is a diagram of the computing power evaluation architecture; DETAILED DESCRIPTION

[0059] like Figure 1 and Figure 2 As shown, the present invention first converts the local pcap file of the training party participating in the training into an image data set through data preprocessing, and then the trusted third-party server sends a lightweight model and data to the training party. The training party feeds back the training time to the trusted third-party server. The trusted third-party server calculates the computing power coefficient based on the training time and decides how many models to divide to the training party according to the computing power coefficient. After the model is divided, the first part is sent to the training party, and the remaining part is sent to the computing server (the server here is only used to run the model and only provides computing resources). After initializing the parameters, training starts from the first training party, and then each training party must obtain parameters from the previous training party before training. One round of training for all training parties is defined as one round, and multiple rounds of training are conducted until the model converges.

[0060] Complete process: Figure 1 System architecture diagram

[0061] ① Data preprocessing stage: defining training data Pid represents the dataset of training party i. Each training party's dataset consists of one or more pcap files of varying lengths. The neural network model requires uniform input length, so the dataset needs to be segmented into files of equal length. Network traffic can be segmented into five types: TCP connection, flow, session, service, and host. A flow is defined as all packets with the same five-tuple (source IP, destination IP, source address, destination address, and transport layer protocol). A flow is unidirectional, while a session is bidirectional. We use the SplitCap tool to segment the pcap files by session, generating several smaller pcap files. We then standardize these smaller pcap files to 784 bytes. Files exceeding 784 bytes are truncated, and any files under 784 bytes are padded with 0x00. This ensures that all data in training party i's dataset Pdi are 784-byte pcap files. These files are then converted into grayscale images in PNG format. Research has shown that using 784 bytes, or 28×28 grayscale images, yields better results than other lengths. Then convert these png images into IDX file types, similar to the MNIST dataset, and become standard input files. After the conversion is completed, the training dataset is

[0062] ②Model distribution stage: / *Training party collection Trusted third-party server Split_sever; server Compute_sever used for model calculation and providing computing resources; * /

[0063] Before distributing the model, the computing power of the training party needs to be evaluated. This patent uses the time required for the training party to complete the model training as the measurement standard. Figure 2 , the trusted third-party server sends a lightweight model and dataset to all trainers, and the trainers send the training time to the third-party server. The trusted third-party server defines the expected training time: τ, the training time of trainer i is Ti, and the computing power coefficient is defined as follows:

[0064]

[0065] When ρ ≥ 1, it means that the average training time of the training party exceeds the expected training time, the computing power is weak, and a hidden layer is allocated to the training party.

[0066] When ρ<1, it means that the average training time of the training party is less than the expected training time, the computing power is strong, and the division hidden layers for the training party, where M represents the number of hidden layers of the model to be trained.

[0067] The third-party server divides the model W according to the computing power coefficient. The first part of the model Wc is sent to the training party, and the remaining model Ws is sent to the computing server. The training party and the computing server cannot know each other's model configuration.

[0068] ③Model training phase: The deep neural network is defined as function F. The neural network has α layers, and the function of running a certain layer is defined as L. Then F(Split_data) is equivalent to L α (L α-1 ...(L1(Split_data))). When backpropagation is performed and the gradient is updated, the gradient update operation from layer α to layer 1 is required. The backpropagation of the entire neural network is defined as F T (gradients), the back propagation of the i-th layer is defined as Then F T (gradient) is equivalent to When forward propagation is performed, the parameters of F() are data. When back propagation is performed, it is to update the gradient, so F T The parameter of () is gradients. The trusted third-party server sends the first β layers to the training party, the β layer is also called the split layer, and the β+1 to α layers are sent to the computing server. Since split learning requires the training party to relay training, the trusted third-party server will randomly sort the training parties to determine the training order. Before training begins, the trusted third-party server randomly generates initialization parameters and sets W C The parameters of Ws are sent to all training parties, and the parameters of Ws are sent to the computing server. It is stipulated that except for the first training party, the other training parties must obtain the model parameters from the previous training party before training. The following is the training phase:

[0069] ① If it is the first training party, it uses the initialization parameters for training. Otherwise, it decrypts the encrypted model parameters sent by the previous training party for training.

[0070] ②Training party i uses local data for training, trains Wc to the last layer, that is, the split layer (β layer) through forward propagation, and sends the fragmented data and labels to the computing server. It can be expressed as:

[0071] L β (L β-1 ...(L1(Sd i )))->F c (Sd i )

[0072] ③ The computing server receives the fragmented data and labels sent by the training party, trains Ws to the last layer through forward propagation, and calculates the loss. It can be expressed as:

[0073] Lα (L α-1 …(L β+1 (F c (Sd i ))))->F S (F c (Sd i ))

[0074] Loss(F S (F c (Sd i )),label)(loss function)

[0075] gradients<-Loss′(F S (F c (Sd i )),label)(gradient)

[0076] (gradients: After calculating the loss function, we need to calculate the gradient of each parameter by taking the partial derivative, and then update the gradient by gradient descent. gradients is just a general term)

[0077] ④ The computing server updates the gradients of each parameter of Ws through back propagation and sends the gradients of the first layer of Ws to the corresponding training party. It can be expressed as:

[0078]

[0079] ⑤ The training party receives the gradient sent by the computing server and updates the gradient of each parameter through back propagation in Wc, which can be expressed as:

[0080] ⑥ After the training party has updated the gradient, it repeats the above process to complete the training of the local data. After the training is completed, the training party establishes a secure channel with the next training party through a trusted third-party server and sends the encrypted model parameters to the next training party.

[0081] ⑦ Repeat the above process until the model converges.

[0082] ⑧ After the model converges, the last trainer sends the model parameters to the trusted third-party server, and the compute server sends the model parameters to the trusted third-party server. The trusted third-party completes the model configuration and broadcasts the completed model to all trainers.

[0083] The training party can filter white traffic after obtaining the model.

[0084] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A white traffic filtering model training method based on multi-party split learning, characterized in that: It includes a trusted third-party server, which is connected to the computing server and several training parties. The specific steps are as follows: Step 1: Data preprocessing: Each training party preprocesses their private data into a unified format; Step 2: Split the white traffic filtering model: The trusted third-party server obtains the average computing power level of all training parties, and then divides the white traffic filtering model into split locations based on the average computing power level of all training parties. The location where the white traffic filtering model is split is used as the split layer. The split layer and the previous white traffic filtering model part are distributed to each training party as the pre-training model, and the remaining part is distributed to the computing server as the post-training model. Step 3: White traffic filtering model training: 3.1) Determine the training order of the training parties; 3.2) The first training party inputs its private data into the pre-training model and performs forward propagation to train the pre-training model. The pre-training model outputs fragmented data, which is then transmitted to the computing server along with the corresponding labels. The computing server inputs the fragmented data and corresponding labels into the post-training model and performs forward propagation to train the post-training model to obtain the output result. The computing server calculates the loss function based on the output results and the labels sent by the first training party. The computing server uses the loss function to perform backpropagation on the post-training model and uses the gradient descent algorithm to update the weights. This continues until it reaches the first layer of the post-training model. The gradient of the first layer of the post-training model is sent to the first training party. The first training party uses the gradient of the first layer of the post-training model to perform backpropagation and update the weights of the first training party's pre-training model to obtain the pre-trained model after training. The first training party transmits the model parameters of the trained previous training model to the next training party; 3.3); Each training party repeats step 3.2) until the white flow filtering model converges; Step 4: The training party and computing server that converge the white traffic filtering model send the pre-training model parameters and post-training model parameters to the trusted third-party server respectively. The trusted third-party server merges the pre-training model and the post-training model to obtain a complete trained white traffic filtering model and broadcasts it to all training parties.

2. The white traffic filtering model training method based on multi-party split learning according to claim 1 is characterized in that: In step 1, the training party's private data Where Pdi represents the dataset of training party i, and N represents the number of training parties. Each training party's dataset contains one or more pcap files, and the dataset is divided into files of equal length.

3. The white traffic filtering model training method based on multi-party split learning according to claim 1 is characterized in that: Use the SplitCap tool to split the pcap file according to the session to generate several small pcap files. The small pcap files are uniformly taken to 784 bytes. For files exceeding 784 bytes, they are intercepted and the ones less than 784 bytes are supplemented with 0x00. Then the small pcap files are converted into 28×28 grayscale images in png format, and then the grayscale images in png format are converted into IDX file type to become standard input files. Pd i It represents the original data set of training party i, Sd i Indicates the dataset after the original training dataset is converted into images.

4. The white traffic filtering model training method based on multi-party split learning according to claim 1 is characterized in that: In step 2, the method for splitting the white traffic filtering model is as follows: The trusted third-party server sends a lightweight model and dataset to all trainers. The trainers send the training completion time to the server. The trusted third-party server defines the expected training time: τ, and the training time of trainer i is T i , the computing power coefficient ρ is defined as follows: When ρ ≥ 1, it means that the average training time of the training party exceeds the expected training time, the computing power is weak, and one hidden layer is allocated to the training party; When ρ<1, it means that the average training time of the training party is less than the expected training time, the computing power is strong, and the training time is divided into 1+ hidden layers for the training party, where M represents the number of hidden layers of the model to be trained; Indicates rounding down (1-ρ)M; The third-party computing power coefficient divides the white traffic filtering model E. The first part of the white traffic filtering model Wc is sent to the training party, and the remaining white traffic filtering model Ws is sent to the computing server, so that the training party and the computing server cannot know each other's model configuration.

5. The white traffic filtering model training method based on multi-party split learning according to claim 4 is characterized in that: In step 3, the white traffic filtering model is a deep neural network, and the deep neural network is defined as a function F. The deep neural network has an α layer, and the function of running a certain layer is defined as L. Then F(Split_data) is equivalent to L α (L α-1 ...(L1(Split_data))), where L α () indicates that forward propagation is performed at layer α; when backpropagation is performed and the gradient is updated, the gradient update operation from layer α to layer 1 is required. The backpropagation of the entire neural network is defined as F T (gradients), the back propagation of the i-th layer is defined as Then F T (gradients) is equivalent to When forward propagation is performed, the parameters of F() are data, gradients represent the gradients of each parameter, and the gradients represent the gradients calculated by the loss function. T represents back propagation; when back propagation is performed, the purpose is to update the gradients, so F F The parameter of () is gradients; the trusted third-party server sends the first β layers to the training party, the β layer is also called the split layer, and the β+1 to α layers are sent to the computing server.

6. The white traffic filtering model training method based on multi-party split learning according to claim 5 is characterized in that: The white traffic filtering model training steps in step 3 are as follows: ① If it is the first training party, it uses the initialization parameters for training. Otherwise, it decrypts the model parameters of the trained model sent by the previous training party for training. ②Training party i uses local data for training, trains Wc to the last layer, that is, the split layer, through forward propagation, and sends the fragmented data and labels to the server, which is expressed as: L β (L β-1 ...(L1(Sd i )))->F c (Sd i ) ③ The server receives the fragmented data and labels sent by the trainer, trains Ws to the last layer through forward propagation, and calculates the loss, which is expressed as: L α (L α-1 …(L β+1 (F c (Sd i ))))->F S (F c (Sd i )) Loss(F S (F c (Sd i )),label) gradients<-Loss′(F S (F c (Sd i )),label) Gradients means that after calculating the loss function, the gradients of each parameter are calculated by taking partial derivatives, and then the gradients are updated by gradient descent. S () indicates that the model performs forward propagation on the computing server side, F c Indicates that the model on the training side performs forward propagation. Loss() represents the loss function, that is, the error between the model's running result and the data's own label. Label represents the data label. Loss′() represents the partial derivative of the loss function to calculate the gradient of the parameter. ④ The computing server updates the gradients of each parameter of Ws through back propagation and sends the gradient of the first layer of Ws to the corresponding training party, which is expressed as: in, Indicates that the back propagation is performed on the computing server model. Indicates that back propagation is performed in the β+1 layer, the gradients of the parameters of the β+1 layer are calculated, and then updated; ⑤ The training party receives the gradient sent by the computing server and updates the gradient of each parameter through back propagation in Wc, which can be expressed as: ⑥ After the training party has updated the gradient, it repeats the above process to complete the training of the local data. After the training is completed, the training party establishes a secure channel with the next training party through a trusted third party and sends the encrypted model parameters of the previous training model to the next training party. ⑦ Repeat steps ① to ⑥ until all training methods have completed the training.

7. A white traffic filtering model training system based on multi-party split learning, characterized in that: The method comprises a computer device configured or programmed to execute the steps of the white flow filtering model training method based on multi-party split learning according to any one of claims 1 to 6.

Images