Industrial internet intrusion detection method based on asynchronous federated learning

Abstract

The invention discloses an industrial internet intrusion detection method based on asynchronous federated learning, and the method comprises the steps: firstly initializing a model, electing a committee leader node, and transmitting the committee leader node to an industrial edge node; each industrial edge node obtains a local model of the industrial edge node after training according to local data of the industrial edge node and uploads the local model; and the committee leader node allocates appropriate weight aggregation to each local model according to the aggregation factor to reform a new global model, circular training is performed until the number of times of training is met, a trained industrial federal intrusion detection model is obtained, and finally to-be-detected industrial internet traffic is input into the model to obtain a detection result. According to the method, the block chain is integrated into industrial federated intrusion detection to deal with the problems of single-point failure, untrusted model construction and the like faced by an industrial federated intrusion detection model, the detection effect and the anti-attack capability are considered, and building of an industrial internet security defense line is facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of Internet security technology and relates to intrusion detection in industrial Internet scenarios, and specifically to an industrial Internet intrusion detection method based on asynchronous federated learning. Background Art

[0002] As an important line of defense in the Industrial Internet, the Intrusion Detection System (IDS) maintains the security of the Industrial Internet by monitoring the entire network to detect attacks. Currently, most intrusion detection schemes related to the Industrial Internet require sufficient high-quality network attack examples for centralized training to build intrusion detection models. However, this is not an easy task in actual industrial scenarios. Generally speaking, a single industrial terminal device usually has very limited attack examples, which makes the construction of intrusion detection models very challenging; at the same time, the owners of industrial terminal devices are usually reluctant to share such highly sensitive attack examples (i.e., those abnormal behavior examples) involving a large number of critical ones with third parties.

[0003] To address this problem, the current mainstream solution is to introduce Federated Learning (FL) into intrusion detection to ensure that the industrial intrusion detection model is built in a privacy-preserving manner. The prior art proposes a federated self-learning anomaly detection system for monitoring the security of IoT devices, which introduces federated learning into anomaly detection. However, the system is limited to a single (Mirai) attack type. The prior art proposes to use the Paillier encryption mechanism to protect the security and privacy of local model parameters, but at the same time, this also generates additional computational overhead, and the model is only applicable to industrial CPS in the same domain. However, the above-mentioned related studies all use a centralized aggregation method for the aggregation of intrusion detection models, which is prone to single point failures, privacy leakage, and unreliable model construction processes.

[0004] To solve the above problems, blockchain technology is considered. As a distributed trusted ledger maintained by multiple parties, blockchain can effectively solve the problems caused by the centralized aggregation of intrusion detection models due to its inherent characteristics of decentralization, traceability, immutability, anonymity, etc. The existing anomaly detection based on blockchain and federated learning uses context-aware Transformer to effectively model the location and time characteristics of vehicle data, and uses a group of reliable consensus node exchange center servers in the alliance blockchain to process all local model updates. However, the above existing industrial federated intrusion detection has the following problems: First, although the current work can achieve effective detection of industrial traffic, there is a problem of unsatisfactory model detection effect caused by the weak ability to extract traffic data features and context perception; secondly, the current work rarely considers the impact of malicious attacks by adversaries on the overall performance of the model during the construction of the industrial federated intrusion detection model.

[0005] Considering the current problems, a new industrial federated intrusion detection method is needed that can take into account both detection effect and anti-attack capability. Summary of the invention

[0006] In view of this, in order to solve the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an industrial Internet intrusion detection method based on asynchronous federated learning (AFL), construct an intrusion detection model based on ACNN-AGRU, and add an aggregation factor to take into account both detection effect and anti-attack capability, thereby solving the problems in the prior art that the industrial federated intrusion detection model has weak ability to extract industrial traffic features and does not pay enough attention to the context information of sequence data, resulting in unsatisfactory detection effect and malicious attacks during the model construction process.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is: In a first aspect, the present application provides an industrial Internet intrusion detection method based on asynchronous federated learning, comprising the following steps: Step 1: The local industrial edge node trains the local model as the initial global model, establishes the genesis block, elects the committee leader node according to the hash value of the genesis block, and initializes the number of training rounds to 0; Step 2: Each industrial edge node downloads the latest global model, trains its own local model based on its own local data and shares it among nodes, and transmits the hash value of the model to the blockchain; Step 3: The committee leader node converts the latest global model (W G L-1 ) and the currently reached local model (W L C ) to generate a new global model W G L , the committee leader node then takes the latest Hash (W G L ) is uploaded to the blockchain to generate a new block; and the new global model W G L Shared with all committee members (industrial edge nodes); Step 4: Use the number of training rounds as the judgment indicator. If it is less than the set number of training rounds required, return to step 2 for the next round of training; otherwise, the loop terminates, and the final intrusion detection model is output and obtained. Step 5: Input the industrial Internet traffic to be tested into the industrial federation intrusion detection model to obtain the intrusion detection result.

[0008] Furthermore, the initial global model in S1 is an intrusion detection model based on ACNN-AGRU, including an attention-based convolutional neural network module, an attention-based gated recurrent unit module, a module composed of a multi-head self-attention mechanism, a multi-layer perceptron module and a Softmax layer.

[0009] Furthermore, in S3, the aggregation factor assigns appropriate weights to the newly arrived local models to weaken the attacks of malicious participants.

[0010] In a second aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the industrial Internet intrusion detection method based on asynchronous federated learning as described in the first aspect above is implemented.

[0011] In a third aspect, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the industrial Internet intrusion detection method based on asynchronous federated learning as described above.

[0012] The beneficial effects of the present invention are: The present invention introduces the channel attention mechanism and the context attention mechanism to improve the intrusion detection model based on ACNN-AGRU, and proposes an aggregation factor to reduce the impact of malicious local models on the global model. The model takes into account both detection effect and anti-attack capability, and solves the problems faced by the industrial federation intrusion detection model, such as single point failure, unreliable model construction, weak ability to extract industrial flow features, and insufficient attention to sequence data context information, resulting in unsatisfactory detection effect. The channel attention mechanism (SEAttention) and contextual attention mechanism (CoTAttention) are introduced into CNN and GRU respectively. On the one hand, it can dynamically adjust the importance of each channel in the output feature map to improve the feature extraction ability of CNN. On the other hand, it can better help GRU capture the contextual information and correlation between sequence data and improve model performance. Based on the aggregation factor, the relationship between the accuracy contribution of the local model and the corresponding industrial terminal reputation value when the global model and the local model are aggregated is comprehensively considered, and appropriate weights are assigned to the local models to ensure that FL can still converge to a higher accuracy safely and efficiently when subjected to poisoning attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0014] Figure 1 This is a flow chart of the intrusion detection method of the present invention; Figure 2 It is a structural schematic diagram of the intrusion detection model of the present invention. DETAILED DESCRIPTION

[0015] The following specific examples are given to further clearly, completely and in detail illustrate the technical solution of the present invention. This example is the best example based on the technical solution of the present invention, but the protection scope of the present invention is not limited to the following examples.

[0016] The industrial control system considered in this invention mainly includes industrial equipment, industrial edge servers and blockchain. The basic idea of ​​AFL-IDS is to network multiple industrial terminal equipment owners (industrial edge nodes with certain computing power) based on the developed industrial federation intrusion detection framework to collaboratively build an intrusion detection model.

[0017] Integrate blockchain into industrial federation intrusion detection to deal with the problems of single point failure and untrustworthy model construction faced by industrial federation intrusion detection models. Aiming at the problem that the current industrial federation intrusion detection model has poor ability to extract industrial traffic features and insufficient attention to the contextual information of sequence data, resulting in unsatisfactory detection results, an intrusion detection model based on ACNN-AGRU is improved. The channel attention mechanism (SEAttention) and the contextual attention mechanism (CoTAttention) are introduced into CNN and GRU respectively. On the one hand, the importance of each channel in the output feature map can be dynamically adjusted to improve the feature extraction ability of CNN, and on the other hand, it can better help GRU capture the contextual information and correlation between sequence data and improve model performance. Then, in order to solve the problem that the industrial intrusion detection model is easily attacked by poisoning by malicious participants during aggregation, an aggregation factor is proposed. The relationship between the accuracy contribution of the local model and the corresponding industrial terminal reputation value when the global model and the local model are aggregated is comprehensively considered, and appropriate weights are assigned to the local model to ensure that FL can still converge to a higher accuracy safely and efficiently when it is attacked by poisoning.

[0018] Example 1 The industrial Internet intrusion detection method based on asynchronous federated learning has the following specific steps: 1. Select an industrial edge node and use its local data to train a local model W L 0 As the initial global model W G 0 , upload the hash value to the blockchain to create the genesis block 0 , according to Block 0 The hash value elects a committee leader node; 2. Each industrial edge node downloads the global model W after receiving the hash value of the latest block G L , and obtain their respective local models (local models) based on their local data training L C , calculate the hash value W of each local model L C And upload them to the blockchain, and at the same time, share the uploaded local models among industrial edge nodes; 3. Without waiting for other nodes, the committee leader node sends the latest global model (W G L-1 ) and the currently reached local model (W L C ) to generate a new global model W G L , and then the committee leader node will send the latest Hash (W G L ) is uploaded to the blockchain to generate a new block. Finally, the new global model W G L Shared with all committee members (industrial edge nodes); 4. Take the number of training rounds as the judgment indicator. If it is less than the set number of training rounds required, return to step 2 for the next round of training; otherwise, the loop terminates, and the final intrusion detection model is output and obtained. 5. Input the industrial Internet traffic to be tested into the industrial federation intrusion detection model to obtain the intrusion detection result. Furthermore, the initial global model W in step 1 above G 0 To improve the intrusion detection model based on ACNN-AGRU, the model structure is as follows Figure 2 As shown, it mainly includes three layers: The first layer includes an attention-based convolutional neural network (ACNN) module, an attention-based gated recurrent unit (AGRU) module, and a module using a multi-head self-attention mechanism; the second layer is a multi-layer perceptron (MLP) module, and the third layer is a Softmax layer.

[0019] Furthermore, in the attention-based gated recurrent unit (AGRU) module, the third layer is a contextual attention mechanism, which specifically includes the following contents: First, all adjacent keys in the K×K grid are spatially represented using K×K group convolutions to represent each key in the context. The learned context key K 1 ∈R H×W×C Reflects the static context information between adjacent keys, K 1 As a static contextual representation of the input X; Afterwards, the context key K 1 and query Q, and then pass two consecutive 1×1 convolutions (W θ With ReLU activation function and W δ Without activation function) to get the attention matrix A: A=[K 1 ,Q]W θ W δ . (1) For each attention head, the local attention matrix A at each spatial location is learned based on the query features and context key features, rather than on independent query-key pairs. 1 Self-attention learning is additionally enhanced. Next, according to the contextual attention matrix A, the participating feature map K is calculated by aggregating all values ​​V 2 ,like: Given the participating feature map K 2 It captures the dynamic feature interactions between inputs, so K 2 The dynamic context representation named as input, and the final output is the static context representation K 1 and the dynamic context representation K 2 fusion of.

[0020] Furthermore, when using federated learning to build an industrial intrusion detection model, the weight of the local model during the aggregation process of the federated learning model has a significant impact on the performance of the federated learning. In order to maintain the high accuracy of the global model when it is attacked by malicious attacks, the committee leader will comprehensively consider the accuracy contribution of its local model and the reputation value of the local node before the model aggregation, and thus design an aggregation factor η, which assigns appropriate weights to the newly arrived local models to weaken the attacks of malicious participants. The specific calculation process of η is as follows: 1.A k i and A G LThey represent the local model accuracy of the current i-th industrial edge node at the k-th aggregation and the current latest global model accuracy, respectively. The accuracy factor a k i The calculation formula is shown in the following formula (3): 2.a k i represents the precision factor of the i-th industrial edge node at the k-th model aggregation, then the reputation value b of the i-th industrial edge node at the k-th aggregation k i The calculation formula is shown in the following formula (4): 3. Aggregation factor η of the i-th industrial edge node at the k-th model aggregation k i is the precision factor a k i and reputation value b k i The product of the two is the aggregation factor η k i The calculation formula is shown in formula (5): 4.W G L-1 is the latest global model before aggregation, W k i is the local model of industrial edge node i at the kth model aggregation. Then, after the kth model aggregation, the latest global model W of industrial edge node i is G L The calculation formula is shown in (6): From formulas (5) and (6), we can intuitively see that if the local node precision factor and reputation value are higher, the corresponding local model aggregation weight factor will be larger, and the global model aggregation effect will be better.

[0021] The above shows and describes the main features, basic principles and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements according to actual conditions, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An industrial Internet intrusion detection method based on asynchronous federated learning, characterized in that: The following steps are involved: S1. The local industrial edge node trains a local model as the initial global model and uploads it to the blockchain to create a genesis block. The committee leader node is elected based on the hash value of the genesis block, where the number of training rounds is initialized to 0. S2. Each industrial edge node downloads the latest global model, trains its own local model based on its own local data and shares it among nodes, and transmits the hash value of the local model to the blockchain; S3. The committee leader node aggregates the latest global model with the currently arrived local model according to the aggregation factor, generates a new global model and shares it among nodes, uploads the hash value of the latest global model to the blockchain, and generates a new block. S4, taking the number of training rounds as the judgment indicator, if it is less than the set number of training rounds required, return to S2 for the next round of training; Otherwise, the loop terminates and the final intrusion detection model is obtained; S5. Input the industrial Internet traffic to be tested into the industrial federation intrusion detection model to obtain the intrusion detection result.

2. The industrial Internet intrusion detection method based on asynchronous federated learning according to claim 1 is characterized in that: The initial global model in S1 is an intrusion detection model based on ACNN-AGRU, including an attention-based convolutional neural network module, an attention-based gated recurrent unit module, a module composed of a multi-head self-attention mechanism, a multi-layer perceptron module and a Softmax layer.

3. The industrial Internet intrusion detection method based on asynchronous federated learning according to claim 2 is characterized in that: The last layer in the attention-based gated recurrent unit module is the contextual attention mechanism.

4. The industrial Internet intrusion detection method based on asynchronous federated learning according to claim 1 is characterized in that: In S3, the aggregation factor assigns appropriate weights to the newly arrived local models to weaken the attacks of malicious participants. The specific calculation process is as follows: S31. Calculate the precision factor of the i-th industrial edge node at the k-th model aggregation S32. Calculate the reputation value of the i-th industrial edge node in the k-th model aggregation S33: Obtain the current aggregation factor based on the product of the above precision factor and the reputation value Then after the kth model aggregation, the latest global model of industrial edge node i for: in, is the latest global model before aggregation, It is the local model of industrial edge node i during the k-th model aggregation.

Images