Cross-domain intelligent monitoring method and system based on byzantine robust privacy protection federated learning
By employing a Byzantine robust privacy-preserving federated learning method, the problems of non-independent identically distributed data and attack threats in intelligent monitoring systems are addressed. This approach achieves normal model convergence and data security, thereby enhancing the system's defense capabilities and privacy protection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Chinese People's Liberation Army Cyberspace Force Information Engineering University
- Filing Date
- 2024-12-13
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional intelligent monitoring systems struggle to converge properly and pose security risks when faced with non-independent, homogeneous monitoring data, privacy attacks, and malware attacks.
We employ a federated learning method based on Byzantine robustness and privacy protection, generate asymmetric key pairs through a key generation center, use the Copaiform Overlapping Community Detection (COPRA) algorithm to divide base station clusters, and design linear homomorphic encryption and robust aggregation strategies to resist poisoning attacks and privacy attacks.
It effectively solves the model convergence problem under non-independent and identically distributed data, enhances data security and system stability, and defends against poisoning attacks and privacy leaks.
Smart Images
Figure CN119903547B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent monitoring technology, and in particular to a cross-domain intelligent monitoring method and system based on Byzantine robust privacy-preserving federated learning. Background Technology
[0002] From 2015 to 2023, the market size of video surveillance equipment in my country doubled, increasing from nearly US$60 billion to US$120 billion, with the growth rate recently remaining at around 8%, indicating a stabilizing market. Urban security, transportation, government projects, finance, and commercial sectors account for over 50% of the market, demonstrating enormous development potential. As a comprehensive technology, video surveillance not only enables video acquisition, processing, control, and emergency command at monitored locations, but also finds widespread application in security, missing persons searches, and social empowerment. In the security field, video surveillance systems effectively monitor and prevent criminal activities, improving the safety of communities and public places. In missing persons searches, video surveillance systems can be used for the search and rescue of missing persons, improving search efficiency and accuracy. Simultaneously, video surveillance technology provides data support and intelligent decision-making for various sectors of society, offering crucial support for urban management, traffic management, and environmental monitoring, thereby empowering society.
[0003] However, traditional intelligent surveillance systems typically rely on centralized artificial intelligence functions located in the cloud or data centers to learn and analyze the identity and behavior of citizens or specific individuals. With the dramatic increase in the amount of video stream data from modern cameras, this centralized solution is inefficient in terms of communication latency, failing to achieve high availability and low cost. This results in high-value surveillance data being discretely distributed across various edge devices, forming data silos. Furthermore, the surveillance data collected by these edge devices often involves privacy issues such as citizens' lifestyles and personal information; leaks could pose significant threats to citizens' personal safety and property. In the future, as the nature of the institutions owning the cameras and the distribution of video data changes, centralized machine learning architectures will no longer be suitable. In this context, federated learning, as a distributed machine learning framework, allows multiple devices to collaboratively train machine learning models without sharing their raw data, making it effective for intelligent surveillance systems. However, facing threats from malicious users and diverse attack methods, while traditional federated learning can solve the "data silo" problem in intelligent surveillance to obtain sufficient datasets to support model accuracy, it cannot ensure data security throughout the entire collaborative modeling process. Based on the aforementioned requirements of the federated learning framework, and considering the practical problems faced by current intelligent monitoring systems in practical applications, this invention aims to conduct a requirements analysis, mainly covering the following aspects:
[0004] (1) The surveillance data collected by cameras deployed in different areas exhibits non-independent and identically distributed characteristics.
[0005] Traditional federated learning assumes that client data is independently and identically distributed (IID). However, in surveillance scenarios, different cameras are deployed in different areas, and the collected surveillance data often exhibits a non-independent and identically distributed (Non-IID) distribution. In this case, the intelligent models trained on different surveillance data show significant differences, leading to difficulties in global model convergence. Therefore, there is an urgent need for a federated learning algorithm capable of handling data heterogeneity to solve the joint modeling challenge of cross-domain intelligent surveillance systems.
[0006] (2) There are advanced privacy attacks and poisoning attacks.
[0007] The adoption of a federated learning architecture in intelligent monitoring systems may expose them to threats of privacy attacks and poisoning attacks, which could severely impact system security and performance. Specifically:
[0008] ① Existing attack methods such as model reverse engineering and model extraction can infer the privacy information of the original training data based on the model parameters transmitted by federated learning. This means that intelligent monitoring based on federated learning still has the risk of privacy leakage.
[0009] ② In the distributed framework of federated learning, any participating node could potentially be a malicious node, uploading samples with misleading data to the federated learning system, such as manipulating image or video data to guide the model to make incorrect identifications or predictions. This could lead to the monitoring system misidentifying specific objects, events, or behaviors, reducing the system's credibility and accuracy. Summary of the Invention
[0010] This invention aims to address the challenges of federated learning failing to converge properly under non-independent and identically distributed data, as well as the threats posed by privacy and poisoning attacks. It proposes a cross-domain intelligent monitoring method and system based on Byzantine robust privacy-preserving federated learning, designs a clustering federated learning architecture to resist the influence of non-independent and identically distributed data, and designs a robust aggregation strategy based on linear homomorphic encryption to effectively resist poisoning and privacy attacks, ensuring data security and confidentiality.
[0011] To achieve the above objectives, the technical solution adopted is:
[0012] A cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning comprises the following five stages:
[0013] Initialization phase: The key generation center generates asymmetric key pairs for the cloud server and each base station. The cloud server generates encrypted initial global model parameters and sends them to all base stations.
[0014] Clustering phase: The cloud server uses the Overlapping Community Detection (COPRA) algorithm to divide base stations with non-independent identically distributed data into multiple clusters with similar data distributions, and selects the leader node in each cluster;
[0015] Intra-cluster security training phase: In each round of training, each base station trains gradients based on local monitoring data, uploads the encrypted gradients to the leader node of the cluster, the leader node identifies and filters poisoned gradients, and then performs synchronous aggregation to obtain intra-cluster encrypted aggregated gradients.
[0016] Global robust aggregation phase: The cloud server first decrypts the encrypted aggregation gradient within the cluster, and then performs weighted aggregation of all aggregation gradients within the cluster using availability weights to obtain the global model parameters for this round of training;
[0017] Real-time monitoring phase: After the model training is completed, each base station obtains a converged intelligent monitoring model, which is then distributed to all cameras within its coverage area for real-time intelligent monitoring.
[0018] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, in the initialization stage, the key generation center generates asymmetric key pairs with linear homomorphic encryption; after receiving the encrypted initial global model parameters sent by the cloud server, the base station decrypts them and performs several rounds of training based on local monitoring data.
[0019] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, in the clustering stage, the cloud server first constructs a weighted undirected graph based on the base station similarity adjacency matrix. In the graph, the nodes represent base stations, and there is a weighted edge between every two nodes. The weight is the similarity of the data distribution of the two base stations. Then, the above weighted undirected graph is used as input, and the COPRA algorithm is used to discover the hidden community structure in the graph. Finally, the base stations are divided into different clusters, and the base stations in each cluster have similar data distributions.
[0020] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, the method for calculating the similarity between base stations is as follows: each base station uploads its respective label vector X to the cloud server. k =(x1,x2,…,x l )∈{0,1} l , where x i =1 indicates that the local dataset of base station k contains label i; the cloud server calculates the similarity between the two base stations according to the Jaccard algorithm.
[0021] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, the selection of the leader node in each cluster further includes: each base station uploading its resource parameters to the cloud server, and the cloud server electing the base station with the strongest computing and communication capabilities from each cluster as the leader node.
[0022] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, the identification and filtering of poisoning gradients by the leader node further includes:
[0023] Given a set of gradients [[G] for cluster K k ]]=g1,g2,...,g n First, the leader node passes through gradient [[g] i Add a random number [r] to the top and mask it for protection, and then add [g] to the top. i [[r]] is sent to the cloud server; then the cloud server receives the ciphertext sent by all the leader nodes, decrypts it using its own private key, calculates the median value, encrypts it, and sends it back to all the leader nodes. The leader nodes subtract [[r]] from the ciphertext to finally obtain the ciphertext of the gradient median value.
[0024] For each cluster's encryption gradient, the leader node selects random number noise as the exponent of the encryption gradient to obtain D. i Simultaneously, the aforementioned random number noise is used as the power exponent of the ciphertext of the gradient median to obtain D. med Then D i and D med Simultaneously, it is sent to the cloud server; the cloud server uses its private key to decrypt and obtain d. i and d med Then calculate d i and d med The cosine similarity is calculated; the cloud server sends the cosine similarity to the leader, who then discards gradients with similarity below a threshold as poisoned gradients.
[0025] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, further, performing synchronous aggregation to obtain intra-cluster encrypted aggregation gradients includes:
[0026] Each base station maintains a cluster list. The cluster label and membership degree are stored in the cluster list in the form of tuples (c, b). The probability that a node belongs to cluster c is represented by the membership degree b. When calculating the encrypted aggregate gradient of cluster k, firstly, the weight of the local gradient of each base station in the cluster is the membership degree of that base station belonging to cluster k divided by the sum of the membership degrees of all base stations in the cluster belonging to cluster k. Then, the leader node performs weighted aggregation of the encrypted gradients of all base stations in cluster k with the weights obtained above to obtain the encrypted aggregate gradient corresponding to cluster k.
[0027] According to the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning of the present invention, the cloud server further performs weighted aggregation of the aggregated gradients within all clusters through availability weights to obtain the global model parameters for this round of training, including:
[0028] The cloud server selects the intra-cluster aggregated gradient with the lowest cosine similarity to the global gradient in the previous round as the benchmark for evaluating poisoned gradients. Then, it calculates the cosine similarity between each intra-cluster aggregated gradient and the benchmark, and calculates the availability weight. The cloud server performs weighted aggregation of all intra-cluster aggregated gradients with availability weight to obtain the global model parameters for this round of training, and distributes them to all base stations. The base stations use the global model parameters to update their local models, thus entering the next round of training.
[0029] Furthermore, this invention also proposes a cross-domain intelligent monitoring system based on Byzantine robust privacy-preserving federated learning, used to implement the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning as described above, including:
[0030] Key Generation Center: Generates asymmetric key pairs for the cloud server and each base station;
[0031] Cameras: collect monitoring data in real time and upload the monitoring data to nearby base stations; cameras receive intelligent monitoring models sent by base stations.
[0032] Base station: Responsible for collecting monitoring data from cameras within its coverage area and, under the coordination of the cloud server, collaborating with other base stations to train intelligent monitoring models.
[0033] Cloud server: Assists base stations in identifying poisoned gradients, is responsible for receiving legitimate gradients submitted by base stations, and performs the final global aggregation.
[0034] The beneficial effects achieved by adopting the above technical solution are:
[0035] 1. This invention designs a clustering federated learning architecture, which effectively solves the problem that federated learning is difficult to converge normally under non-independent and identically distributed data.
[0036] This invention abstracts the similarity between base station data distributions in federated learning into a network structure. Nodes are the base stations participating in federated learning, collecting surveillance data from cameras within their coverage area; edges are the similar links between base stations; and the hidden community structure between nodes represents the cluster structure in clustering federated learning. This invention introduces the classic overlapping community detection algorithm COPRA, which can discover potential clustering structures between base stations, grouping base stations with similar data distributions into one class. This effectively solves the convergence problem of federated learning and also aids in the detection of poisoning attacks.
[0037] 2. This invention designs a robust aggregation strategy based on linear homomorphic encryption, which effectively resists poisoning attacks and privacy attacks, and ensures data security and confidentiality.
[0038] To enhance the security and stability of the intelligent monitoring system, a trusted key generation center generates different key pairs for the cloud server and base stations to ensure communication security. Each base station is trained locally to obtain gradients, which are then converted into ciphertext using the Paillier homomorphic encryption algorithm. Upon receiving the encrypted gradients, the leader, in conjunction with the cloud server, executes a secure median algorithm, a secure cosine similarity algorithm, and an intra-cluster weighted aggregation algorithm to detect potentially malicious gradients and complete secure intra-cluster aggregation. This process effectively resists poisoning attacks and guarantees data privacy. Attached Figure Description
[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.
[0040] Figure 1 This is one of the flowcharts of the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning, an embodiment of the present invention.
[0041] Figure 2 This is the second flowchart of the cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning in an embodiment of the present invention.
[0042] Figure 3 It is a weighted undirected graph according to an embodiment of the present invention. Detailed Implementation
[0043] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.
[0044] like Figure 1 and Figure 2 As shown in the figure, this embodiment discloses a cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning, which includes five stages: initialization stage, clustering stage, intra-cluster secure training stage, global robust aggregation stage, and real-time monitoring stage. Among them, only the intra-cluster secure training stage and the global robust aggregation stage need to be repeated in each round of training until a trained intelligent monitoring model is obtained.
[0045] Step S101, Initialization phase.
[0046] During the initialization phase, a trusted key generation center generates a linear homomorphic encrypted asymmetric key pair (pk) for the cloud server. s ,sk s ), and at the same time for each base station U i Generate asymmetric key pairs (pk i ,sk i The cloud server generates encrypted initial global model parameters [[W (0) The data is then sent to all base stations, which decrypt it and perform several rounds of training based on their local monitoring data.
[0047] Step S102, Clustering Stage.
[0048] During the clustering phase, the cloud server uses the Overlapping Community Detection Algorithm (COPRA) to divide base stations with non-independent identically distributed data into multiple clusters with similar data distributions, and selects the leader node in each cluster.
[0049] Observations reveal that the clustering federated learning architecture resembles the community structure in complex networks. Complex networks are typically composed of various community structures, with nodes within the same community being tightly connected and nodes in different communities being sparsely connected. If we abstract the similarity of data distribution among the base stations participating in federated learning as a network structure, then nodes are the base stations participating in federated learning, edges are the similar links between base stations, and the hidden community structure between nodes is the cluster structure in clustering federated learning. This scheme introduces the classic overlapping community detection algorithm COPRA, which can discover potential cluster structures among base stations and achieve dynamic and flexible clustering without pre-defining the number of clusters.
[0050] Based on these ideas, this proposal suggests a similarity-based soft clustering federated learning algorithm. Let S... U Let U be a set of base stations and L be a set of l tags. Then, in the case of Non-IID data distribution, each base station has a subset of tags from set L. Define base stations k∈S. U The label vector is X k =(x1,x2,…,x l )∈{0,1} l , where x i =1 indicates that label i exists in the local dataset of base station k. To effectively calculate the similarity of data distribution among base stations and meet real-time requirements, the focus is on whether a certain type of label exists in the local monitoring dataset of each base station, thereby obtaining the label vector of the base station. The similarity between base stations is measured by evaluating the similarity of these label vectors. First, each base station uploads its respective label vector X to the server. k The server then calculates the similarity between each pair of base stations, obtaining a similarity matrix. Since the label vectors are binary vectors, the Jaccard algorithm is chosen to calculate the similarity between vectors because it only considers elements with a value of 1, not elements with a value of 0. The formula for calculating similarity using the Jaccard algorithm is: J(A,B)=|A∩B| / |A∪B|, where A and B represent two vectors, |A∩B| represents the number of matching labels in vectors A and B, and |A∪B| represents the total number of labels with a value of 1 in both label distribution vectors. Therefore, the similarity between two base stations can be calculated as:
[0051] The cloud server constructs a weighted undirected graph based on the base station similarity adjacency matrix. Nodes in the graph represent base stations, and every two nodes share a weighted edge. The weight of each edge represents the similarity in the data distribution between the two base stations. This weighted undirected graph is then used as input, and the COPRA algorithm is employed to discover the hidden community structure within the graph, which represents the clustering structure between base stations.
[0052] However, in practice, when the similarity is below a certain threshold, the probability that two nodes belong to the same community is very low. The COPRA algorithm, when updating node labels, must traverse the label lists of all neighboring nodes. Therefore, to reduce the computational overhead of label propagation, some edges with low similarity are removed to improve the efficiency of community detection. For example, suppose there are 10 base stations participating in this round of federated learning training. Edge weights are visualized through edge length: the shorter the edge, the higher the similarity between the data distributions of different base stations. The weighted undirected graph before deletion is shown below. Figure 3 As shown in (a), there is a similar link between every two base station nodes. The weighted undirected graph after deletion is as follows. Figure 3 As shown in (b), only a portion of the highly similar links were retained. It is evident that in... Figure 3 (a) and Figure 3 In (b), three distinct cluster structures exist. Therefore, deleting links with similarity below a certain threshold will not significantly affect the final community detection results. Furthermore, due to the reduction in similar links between nodes, the community structure becomes clearer. In practical applications, deleting redundant similar links can not only reduce the computational cost of constructing a weighted undirected graph, but also reduce the computational cost of cluster label propagation while ensuring the accuracy of cluster detection results, thus improving the efficiency of cluster label propagation.
[0053] Through the clustering process described above, base stations are divided into different clusters. Base stations within each cluster have similar data distributions, thus mitigating the impact of non-independent and identically distributed data. Simultaneously, each base station must upload its resource parameters V (e.g., bandwidth and computing power) to the cloud server. The cloud server then elects the base station with the strongest computing and communication capabilities from each cluster k as the leader node LN. kIt is responsible for the next stage of cluster aggregation and poisoning attack detection.
[0054] Step S103, Intra-cluster security training phase.
[0055] In each round of training, each base station U i Gradients were trained based on local monitoring data. After encryption using Paillier, the gradients are uploaded to the leader node, which identifies and filters the poisoned gradients, then performs synchronous aggregation to obtain the intra-cluster encrypted aggregated gradients. This stage specifically includes the following three steps:
[0056] Step ①: Design the Smed function to compute the median gradient without revealing any privacy. The pseudocode for the Smed function is as follows. Specifically, given a set of gradients [[G] of cluster K... k ]]=g1,g2,...,g n First, the leader node LN k By gradient [[g i Add a random number [r] to the top and mask it for protection, and then add [g] to the top. i The ciphertext is sent to the cloud server. Then, after the cloud server receives the ciphertext sent by all the leader nodes, it decrypts it using its own private key, calculates the median value, encrypts it, and sends it back to all the leader nodes. The leader nodes subtract the ciphertext of the random number r from the ciphertext to finally obtain the ciphertext of the gradient median value.
[0057]
[0058]
[0059] Step 2: Design the SCos function to calculate the cosine similarity of the encryption gradients while protecting privacy. The pseudocode for the SCos function is as follows. Specifically, for each intra-cluster encryption gradient... Leader node selection random number noise r i∈[1,n] take it as The power of the result is used to obtain D. i At the same time, r i∈[1,n] Ciphertext as the median of the gradient The power of the result is used to obtain D. med Then D i and D med It is also sent to the cloud server. The cloud server uses its private key to decrypt and obtain d. i and d med Then calculate d i and d med The cosine similarity, based on the properties of additive homomorphic encryption, involves random number noise r when calculating this cosine similarity. i∈[1,n]It is eliminated, therefore its result is equal to the intra-cluster gradient. Compared with the median value The cosine similarity is calculated. The cloud server sends the cosine similarity to the leader, which discards gradients with similarity below a threshold as poisoned gradients.
[0060]
[0061]
[0062] Step 3: The COPRA algorithm described above can obtain the clustering structure between base stations and the label of each class. However, some base stations may belong to multiple classes, and their membership degree to each class is different. Therefore, each base station maintains a cluster list, and the cluster label and membership degree are stored in the cluster list in the form of tuples (c, b). The probability of a node belonging to cluster c is represented by the membership degree b. The membership degree represents the degree of preference of the base station for different clusters. From the perspective of cluster federated learning, base stations with different membership degrees contribute differently to the intra-cluster model update. The higher the membership degree, the greater the contribution of the base station weight to the intra-cluster model update. Therefore, this scheme adopts a weighted average strategy for intra-cluster aggregation, and the weight is related to the membership degree of the base station within its cluster. For example, in a certain round of training, the cluster list of the base station is updated. At this time, cluster k contains base stations 1, 2, and 3, and their cluster label lists are respectively When calculating the encryption aggregation gradient of cluster k, the weights of the local gradients of base stations 1, 2, and 3 are calculated as follows:
[0063]
[0064] Then, the leader node LN k Based on the additive homomorphism of the Paillier algorithm, the encryption gradients of the three base stations are weighted and aggregated according to the weights obtained above to obtain the encryption aggregated gradient corresponding to cluster k:
[0065]
[0066] Step S104, Global Robust Aggregation Phase.
[0067] Considering more complex attack scenarios, attackers might compromise the leader node to disrupt the training process within the cluster, or there might be more than 50% malicious nodes in the system. In such cases, the contribution of each cluster's aggregated gradient to the global model would be weakened. Therefore, this solution designs a globally robust aggregation algorithm. The basic idea is as follows: First, the cloud server decrypts the encrypted aggregated gradient within the cluster. Then, the cloud server selects the gradient g from the previous round... w The intra-cluster aggregation gradient with the lowest cosine similarity (note that this gradient is the decrypted gradient) is used as the benchmark g0 for evaluating the poisoning gradient. Then, the intra-cluster aggregation gradient g is calculated for each cluster.k The cosine similarity to the benchmark is calculated, and usability weights are set as follows:
[0068]
[0069] It is evident that the closer the cluster-level aggregated gradient is to the poisoning baseline, the lower its availability. Finally, the cloud server performs weighted aggregation of all cluster-level aggregated gradients using availability weights to obtain the global model parameters for this round of training (as shown in the following formula), and distributes them to all leader nodes. The leader nodes distribute these global model parameters to all base stations within their clusters, and the base stations use these global model parameters to update their local models, thus entering the next round of training.
[0070]
[0071] w t ←w t-1 -ηg w
[0072] Step S105, Real-time monitoring stage.
[0073] After the federated learning process, which includes the initialization phase, clustering phase, intra-cluster security training phase, and global robust aggregation phase, each base station ultimately obtains a converged intelligent monitoring model. This model is then distributed to all intelligent monitoring cameras within its jurisdiction. The cameras can then utilize this model for real-time intelligent monitoring, enabling tasks such as intelligent security and pedestrian recognition.
[0074] To realize the aforementioned cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning, this embodiment also discloses a cross-domain intelligent monitoring system based on Byzantine robust privacy-preserving federated learning, including several cameras and base stations, a key generation center and a cloud server.
[0075] Key Generation Center: An independent and trusted entity that generates and distributes public and private key pairs for cloud servers and base stations. These keys are used to encrypt the local gradients of the base stations, thereby preventing the leakage of any sensitive information from local data.
[0076] Cameras: collect monitoring data in real time and upload the data to nearby base stations; cameras have AI functions and can receive intelligent monitoring models sent by base stations to achieve tasks such as intelligent security and pedestrian recognition.
[0077] Base stations are responsible for collecting surveillance data from cameras within their coverage area and, under the coordination of cloud servers, collaborating with other base stations to train intelligent monitoring models. Base stations can be benign or malicious; malicious base stations will submit encrypted, toxic gradients to corrupt the global model.
[0078] Cloud server: Assists base stations in identifying poisoned gradients, is responsible for receiving legitimate gradients submitted by base stations, and performs the final global aggregation. The cloud server is considered semi-honest, meaning it will honestly perform all operations according to the protocol while striving to obtain more sensitive information about the monitoring data.
[0079] In summary, this invention possesses excellent defense precision. It can simultaneously defend against poisoning attacks in both IID and Non-IID data environments, ensuring the global accuracy of the intelligent monitoring model remains within a reasonable range. Furthermore, this invention offers strong privacy protection. It guarantees the confidentiality of gradient parameters, preventing any attacker from stealing the privacy of the monitoring data.
[0080] Unless otherwise specifically stated, the components, steps, numerical expressions, and values set forth in these embodiments do not limit the scope of the invention.
[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0082] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.
[0083] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.
[0084] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features within the technical scope disclosed in the present invention; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning, characterized in that, It includes the following five stages: Initialization phase: The key generation center generates asymmetric key pairs for the cloud server and each base station. The cloud server generates encrypted initial global model parameters and sends them to all base stations. Clustering phase: The cloud server uses the Overlapping Community Detection (COPRA) algorithm to divide base stations with non-independent identically distributed data into multiple clusters with similar data distributions, and selects the leader node in each cluster; Intra-cluster security training phase: In each round of training, each base station trains gradients based on local monitoring data, uploads the encrypted gradients to the leader node of the cluster, the leader node identifies and filters poisoned gradients, and then performs synchronous aggregation to obtain intra-cluster encrypted aggregated gradients. Global Robust Aggregation Phase: The cloud server first decrypts the encrypted aggregated gradients within the cluster. The cloud server selects the aggregated gradient within the cluster with the lowest cosine similarity to the global gradient from the previous round as the benchmark for evaluating poisoned gradients. Then, it calculates the cosine similarity between each aggregated gradient within the cluster and the benchmark, and calculates the availability weight. Finally, the cloud server performs weighted aggregation of all aggregated gradients within the cluster using the availability weight to obtain the global model parameters for this round of training, and distributes them to all base stations. The base stations use these global model parameters to update their local models, thus entering the next round of training. Real-time monitoring phase: After the model training is completed, each base station obtains a converged intelligent monitoring model, which is then distributed to all cameras within its coverage area for real-time intelligent monitoring.
2. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 1, characterized in that, During the initialization phase, the key generation center generates asymmetric key pairs for linear homomorphic encryption; after receiving the encrypted initial global model parameters sent by the cloud server, the base station decrypts them and performs several rounds of training based on local monitoring data.
3. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 1, characterized in that, In the clustering stage, the cloud server first constructs a weighted undirected graph based on the base station similarity adjacency matrix. In the graph, nodes represent base stations, and there is a weighted edge between every two nodes. The weight is the similarity of the data distribution of the two base stations. Then, the above weighted undirected graph is used as input, and the COPRA algorithm is used to discover the hidden community structure in the graph. Finally, the base stations are divided into different clusters, and the base stations in each cluster have similar data distributions.
4. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 3, characterized in that, The method for calculating the similarity between base stations is: Let Let U be a set of U base stations, and L be a set of l tags. Each base station has a subset of tags from set L. Upload their respective tag vectors to the cloud server ,in This indicates that the local dataset of base station k contains label i; the cloud server calculates the similarity between the two base stations using the Jaccard algorithm.
5. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 1, characterized in that, The process of selecting a leader node in each cluster involves each base station uploading its resource parameters to the cloud server, and the cloud server electing the base station with the strongest computing and communication capabilities from each cluster as the leader node.
6. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 1, characterized in that, The poisoning gradients identified and filtered by the leader node include: Given a set of gradients for cluster K First, the leader node passes through the gradient. Add random number Protect it with a mask and The ciphertext is sent to the cloud server; then, after receiving the ciphertext from all the leader nodes, the cloud server decrypts it using its own private key, calculates the median value, encrypts it again, and sends it back to all leader nodes. The leader nodes subtract [a certain value] from the ciphertext. Finally, the ciphertext of the gradient median value is obtained; For each cluster's encryption gradient, the leader node selects random number noise as the exponent for the encryption gradient calculation. Simultaneously, the aforementioned random number noise is used as the power of the ciphertext of the gradient median value for calculation. Then and Simultaneously, it is sent to the cloud server; the cloud server uses its private key to decrypt and obtain the data. and Then calculate and The cosine similarity is calculated; the cloud server sends the cosine similarity to the leader, who then discards gradients with similarity below a threshold as poisoned gradients.
7. The cross-domain intelligent monitoring method based on Byzantine robust privacy-preserving federated learning according to claim 1, characterized in that, The intra-cluster encrypted aggregation gradient obtained by performing synchronous aggregation includes: Each base station maintains a cluster list. The cluster label and membership degree are stored in the cluster list in the form of tuples (c, b). The probability that a node belongs to cluster c is represented by the membership degree b. When calculating the encrypted aggregate gradient of cluster k, firstly, the weight of the local gradient of each base station in the cluster is the membership degree of that base station belonging to cluster k divided by the sum of the membership degrees of all base stations in the cluster belonging to cluster k. Then, the leader node performs weighted aggregation of the encrypted gradients of all base stations in cluster k with the weights obtained above to obtain the encrypted aggregate gradient corresponding to cluster k.
8. A cross-domain intelligent monitoring system based on Byzantine robust privacy-preserving federated learning, characterized in that, A method for implementing cross-domain intelligent monitoring based on Byzantine robust privacy-preserving federated learning as described in any one of claims 1-7, comprising: Key Generation Center: Generates asymmetric key pairs for the cloud server and each base station; Cameras: collect monitoring data in real time and upload the monitoring data to nearby base stations; cameras receive intelligent monitoring models sent by base stations. Base station: responsible for collecting monitoring data from cameras within its coverage area, and, under the coordination of the cloud server, collaborating with other base stations to train intelligent monitoring models; Cloud server: Assists base stations in identifying poisoned gradients, is responsible for receiving legitimate gradients submitted by base stations, and performs the final global aggregation.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.