Hierarchical clustering method for realizing Byzantine robustness and privacy protection in federated social network

Abstract

The invention provides a Byzantine robust and private hierarchical clustering method in a federated social network, which comprises the following steps of: firstly, carrying out random response disturbance processing on adjacent vectors of users in the social network, and then authenticating the adjacent vectors with noise by using lightweight random masks; and finally, iteratively learning the optimal hierarchical clustering tree by using a coarse-grained quality function, and applying the generated optimal hierarchical clustering tree to a social recommendation system. According to the method, the optimal hierarchical clustering tree of the federated social network can be safely and privately learned, edge local differential privacy protection is provided, the Byzantine robustness is ensured, and only one-time interaction with the user is carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of social network privacy protection, and particularly to a method for implementing Byzantine-robust and privacy-protected hierarchical clustering in a federated social network. Background Art

[0002] For years, people have been researching hierarchical clustering to discover relevant information and predict user behavior in social networks. At the same time, due to data privacy issues in centralized networks, distributed services have gradually gained favor among users. For example, Meta is currently developing a text-based decentralized social network platform, which is expected to become a strong competitor to Twitter. This trend towards decentralization has prompted the development of hierarchical clustering towards federated social networks, which provides strong support for many end applications such as personalized recommendations and online advertising.

[0003] However, this new paradigm, namely hierarchical clustering in a federated social network, poses the problem of redesigning Byzantine-robust and privacy-protected hierarchical clustering algorithms for the federated setting, aiming to achieve data privacy and Byzantine robustness. Specifically, in this new paradigm, while users query the terminal service, they keep the privacy of their social relationships (i.e., adjacency vectors), thus maintaining the privacy of the entire social relationship. The centralized server cannot use the network structure in this federated setting. In contrast, most existing work is centralized, and the untrusted centralized server can obtain the structure of this network, which leaks the data privacy of users. At the same time, they do not consider Byzantine robustness either, because malicious users (i.e., Byzantines) may deliberately poison the data in the federated setting. Therefore, the related research work cannot guarantee the privacy and Byzantine robustness of users in social networks. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for implementing Byzantine-robust and privacy-protected hierarchical clustering in a federated social network, which can reduce the risk of privacy leakage of users in social networks and provide Byzantine robustness at the same time.

[0005] The technical solution adopted by the present invention to solve its technical problems is: a method for implementing Byzantine-robust and privacy-protected hierarchical clustering in a federated social network, including the following steps:

[0006] (1) Local differential perturbation with adaptive noise;

[0007] (2) Authentication of lightweight random masks;

[0008] (3) Learning the optimal hierarchical clustering tree using a mixed quality function;

[0009] (4) Apply the optimal hierarchical clustering tree to the social recommendation system.

[0010] Specifically, step (1) is as follows: Use the randomized response mechanism to perform local perturbation processing on the adjacency vector of each user, and use a specially designed noise adaptive adjustment method based on the graph structure to reduce the noise in the adjacency matrix. By optimizing the formula Output the optimized adjacency matrix, where is the perturbed adjacency matrix, and A c is the optimized adjacency matrix, and λ is the optimization learning rate;

[0011] Specifically, step (2) is as follows: In the user registration stage, use a carefully designed two-party secret sharing mechanism, and each user registers its adjacency vector on the server; in the authentication stage, design a lightweight two-party secure inner product protocol so that the server can verify the perturbed data without decrypting the data;

[0012] In step (3), the optimal hierarchical clustering tree is obtained by the user sending the noisy adjacency vector to the server at one time. The server first performs preliminary clustering based on the coarse-grained quality function, and then optimizes the clustering result using the fine-grained quality function.

[0013] Specifically, step (3) is as follows: First, based on the optimized adjacency matrix A c , construct the coarse-grained quality function Q D (T) = ∑ x∈v,y∈v C(x, y)|leaves(T[x ∨ y])|, where C(x, y) is the similarity matrix composed of the cosine distances between any two nodes x and y in the social network belonging to the clusters, and T[x ∨ y] is the number of leaf nodes of the smallest subtree containing x and y;

[0014] Step (3) also includes: Using the Metropolis-Hastings (MH) algorithm, based on the Markov chain as the initial state, sample an initial hierarchical clustering tree T 0 , in the iterative process, randomly select an internal node λ i , and generate two alternative configurations based on the current configuration; for each alternative configuration, calculate its corresponding coarse-grained quality function Q D (T i+1 ); then calculate the probability β of T i+1 based on the MH algorithm, where According to the probability β, decide whether to accept the new state T i+1 as the hierarchical clustering tree for the next iteration, or keep the current state unchanged.

[0015] Step (4) is specifically as follows: In both the existing user and cold start modes, the optimal hierarchical clustering tree generated in step (3) is extended to the social recommendation algorithm, and a movie recommendation system is designed simultaneously.

[0016] Advantageous Effects

[0017] Due to the adoption of the above technical solutions, compared with the prior art, the present invention has the following advantages and positive effects: By exploring hierarchical clustering in the social network, randomly perturbing the adjacent vectors of users in the social network, and using a noise-adaptive graph structure to reduce noise and generate optimized adjacent vectors; performing lightweight random masking authentication on the optimized adjacent vectors to ensure Byzantine robustness; using a hybrid quality function to learn the optimal hierarchical clustering tree; and theoretically combining differential privacy to prove the privacy protection performance boundary and Byzantine robustness of this mechanism. The present invention breaks through the limitations of the prior art research on privacy protection and Byzantine robustness in the hierarchical clustering of federated networks, and for the first time pays attention to the poor noise performance and Byzantine robustness in the hierarchical clustering of federated networks. It innovatively proposes to generate optimized adjacent vectors by reducing noise according to a noise-adaptive graph structure and a two-stage lightweight authentication scheme to ensure Byzantine robustness, providing an important supplement to the research on privacy protection and Byzantine robustness technologies for hierarchical clustering, with the expectation of further reducing the risk of user privacy leakage in hierarchical clustering, avoiding malicious users from poisoning the data, and promoting the continuous enrichment and improvement of the relevant theoretical systems and technologies for hierarchical clustering privacy protection and Byzantine robustness. Description of the Drawings

[0018] Figure 1 It is a schematic diagram of how a social network graph is converted into an adjacency matrix;

[0019] Figure 2 It is a schematic flowchart of the method for implementing Byzantine-robust and privacy-protected hierarchical clustering in a federated network in the embodiment;

[0020] Figure 3 It is a general research idea diagram of the method for implementing Byzantine-robust and privacy-protected hierarchical clustering in a federated network in the embodiment;

[0021] Figure 4 It is a schematic flowchart of ensuring Byzantine robustness in the federated social network in the embodiment;

[0022] Figure 5 It is a schematic structural diagram of two alternative configurations generated based on the current configuration in the embodiment;

[0023] Figure 6 It is a schematic diagram of applying the optimal hierarchical clustering tree generated in the embodiment to a movie recommendation system. Detailed Embodiments

[0024] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0025] Figure 1 shows how a social network graph of the present invention is converted into an adjacency matrix schematic diagram. As Figure 1 shown, each user in the social network has its own corresponding adjacency vector. An element of 1 in the vector represents that two users are friends, and 0 represents non-friends.

[0026] Figure 2 shows a schematic flowchart of a method for implementing Byzantine-robust and privacy-preserving hierarchical clustering in a federated network. In this embodiment, the processing process of all possible users in the social network is taken as an example for illustration.

[0027] As Figure 2 shown, the method for implementing Byzantine-robust and privacy-preserving hierarchical clustering in this embodiment includes:

[0028] Step S301: Trusted users perform local random perturbation, and malicious users perform data poisoning;

[0029] Step S302: All users send authentication requests;

[0030] Step S303: The semi-trusted server authenticates the adjacency vectors;

[0031] Step S304: The semi-trusted server sends the authentication results;

[0032] Step S305: All users send the perturbed adjacency vectors;

[0033] Step S306: The semi-trusted server iteratively learns the optimal hierarchical clustering tree.

[0034] According to the solution of this embodiment, random response perturbation is performed on the adjacency vectors of trusted users in the social network, and a graph structure based on noise adaptation is used to reduce noise to generate optimized adjacency vectors. Malicious users may perform data poisoning; lightweight random masking authentication is performed on the optimized adjacency vectors to identify malicious users in the social network, thereby ensuring Byzantine robustness; a hybrid quality function is used to iteratively learn the optimal hierarchical clustering tree; and the privacy protection performance boundary and Byzantine robustness of this mechanism are proven theoretically in combination with differential privacy.

[0035] Based on the above-described embodiments, Figure 3The flowchart shows a method for achieving Byzantine-robust and privacy-preserving hierarchical clustering in a federated network in a specific example.

[0036] In this example, the trusted users send the perturbed adjacency vectors to the semi-trusted server after random response perturbation locally. However, the noise in the perturbed adjacency vectors is too large. Then the semi-trusted server uses the graph structure method based on adaptive noise to reduce the noise in the adjacency matrix, by optimizing the formula to output the optimized adjacency matrix. This example aims to further reduce the risk of user privacy leakage in hierarchical clustering and promote the continuous enrichment and improvement of the relevant theoretical system and technology of hierarchical clustering privacy protection.

[0037] Figure 4 The flowchart shows a method for ensuring Byzantine robustness in a federated social network in an example. To ensure Byzantine robustness, the present invention needs to verify the data of users in the social network because Byzantines can easily poison the data in the federated network. Therefore, the present invention designs a two-stage lightweight random masking authentication method to ensure the Byzantine robustness of the adjacency vectors. In the first stage, users register the adjacency vectors based on a carefully designed two-party secret sharing; in the second stage, namely similarity authentication, the server authenticates the reported data through a two-party secure inner product protocol. This example aims to avoid poisoning attacks by malicious users in the federated social network and promote the continuous enrichment and improvement of the relevant theoretical system and technology of hierarchical clustering Byzantine robustness.

[0038] Figure 5 The structural diagram shows two alternative configurations generated based on the current configuration in an embodiment. To obtain the optimal hierarchical clustering tree, first, based on the optimized adjacency matrix A c , construct the coarse-grained quality function Q D (T) = ∑ x∈v,y∈v C(x, y)|leaves(T[x ∨ y])|; Use the Metropolis-Hastings (MH) algorithm, based on the Markov chain as the initial state, sample an initial hierarchical clustering tree T 0 from the current distribution. During the iteration process, randomly select an internal node λ i , and generate two alternative configurations based on the current configuration, as shown in Figure 5 ; For each alternative configuration, calculate its corresponding coarse-grained quality function Q D (T i+1) ; Then calculate the probability β of T i+1 based on the MH algorithm. According to the probability β, decide whether to accept the new state T i+1 as the hierarchical clustering tree for the next iteration or keep the current state unchanged.

[0039] Figure 6 The figure shows the optimal hierarchical clustering tree generated in an embodiment applied to a movie recommendation system. In both the existing user and cold start modes, the generated optimal hierarchical clustering tree is extended to a social recommendation algorithm, and a movie recommendation system is designed for recommendation, as Figure 6 shown.

Claims

1. A method for implementing Byzantine robust and private hierarchical clustering in federated social networks, characterized in that The optimal hierarchical clustering tree is generated by constructing a coarse-grained quality function, including the following steps: (1) Local differential perturbation with adaptive noise; (2) Lightweight random mask authentication; (3) Using a mixed quality function to learn the optimal hierarchical clustering tree; (4) Apply the optimal hierarchical clustering tree to social recommendation system.

2. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The step (1) is specifically as follows: a random response mechanism is used to perform local disturbance processing on the adjacency vector of each user, and a specially designed graph-structure-based noise adaptive adjustment method is used to reduce the noise in the adjacency matrix. By optimizing the formula Output the optimized adjacency matrix, where is the perturbed adjacency matrix, A c is the optimized adjacency matrix, and λ is the optimized learning rate.

3. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The step (2) is specifically as follows: in the user registration stage, each user registers his / her adjacency vector on the server using a carefully designed two-party secret sharing mechanism; In the authentication phase, a lightweight two-party secure inner product protocol is designed to enable the server to verify the perturbed data without decrypting the data.

4. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The optimal hierarchical clustering tree in step (3) is obtained by sending the noisy adjacency vectors to the server at one time, the server first performs preliminary clustering based on a coarse-grained quality function, and then optimizes the clustering results using a fine-grained quality function.

5. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The step (3) is specifically as follows: first, based on the optimized adjacency matrix A c , construct the coarse-grained quality function Q of the hierarchical clustering tree D (T) = ∑ x∈v，y∈v C(x,y)|leaves(T[x∨y])|, where C(x,y) is the similarity matrix formed by the cosine distance between the clusters to which any two nodes x and y belong in the social network, and T[x∨y] is the number of leaf nodes in the smallest subtree containing x and y.

6. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The step (3) further includes: using the Metropolis-Hastings (MH) algorithm, based on the Markov chain as the initial state, sampling from the current distribution to obtain an initial hierarchical clustering tree T0, and randomly selecting an internal node λ during the iteration process. i , and generate two alternative configurations based on the current configuration; for each alternative configuration, calculate its corresponding coarse-grained quality function Q D (T i+1 ), and then calculate T based on the MH algorithm i+1 The probability β, where According to the probability β, the decision is to accept the new state T i+1 As the hierarchical clustering tree for the next iteration, the current state remains unchanged.

7. The method for implementing Byzantine robust and private hierarchical clustering in a federated social network according to claim 1, characterized in that: The step (4) is specifically as follows: in both the existing user mode and the cold start mode, the optimal hierarchical clustering tree generated in the step (3) is used to expand the social recommendation algorithm, and a movie recommendation system is designed.

Images