Social vehicle-mounted network privacy protection routing solution based on local differential privacy

Abstract

The invention designs and invents a social vehicle-mounted network privacy protection routing solution based on local differential privacy. The method is mainly characterized by comprising the following steps: (1) calculating social attributes of users based on a clustering algorithm of AP features; (2) a hotspot position confusion algorithm based on local differential privacy; (3) dividing vehicle communities based on a spectral clustering algorithm and vehicle hotspot positions; and (4) integrating a message forwarding strategy of the multi-community characteristics of the user. From the sociality of nodes and the privacy of social attributes, the invention provides a social vehicle-mounted network privacy protection routing solution based on local differential privacy, and aims to improve the security performance and the transmission performance of a social vehicle-mounted network, calculate the hotspot position of a user based on an AP clustering algorithm, determine the social attributes of a vehicle, and improve the security and the transmission performance of the social vehicle-mounted network. According to the method, vehicle hotspot positions are confused through a local differential privacy algorithm, the privacy of a user is protected, meanwhile, the availability of a social vehicle-mounted network is ensured, the confused hotspot positions are clustered, and communities are allocated to vehicles. In the message transmission process, the community of the vehicle and the activeness of the vehicle are considered, the message delivery rate is improved, the message transmission hop count is reduced, the number of copies is controlled in the message transmission process, and waste of network resources caused by excessive copies is avoided.

Description

Technical field:

[0001] The present invention relates to the fields of privacy protection and wireless communication technology, and is a solution for privacy protection routing in a community opportunity network. Background technology:

[0002] Opportunistic networks, composed of wireless communication devices carried by vehicles or people, exchange messages by leveraging opportunities when nodes encounter each other during movement. They are a research hotspot in the communications field. Opportunistic networks are characterized by frequent node movement, unstable message pathways, and real-time changes in network structure. Exploiting these characteristics and improving their performance is a challenge in opportunistic network communication research. Given that opportunistic networks are based on people, nodes often form strong relationships based on shared social characteristics such as interests, hobbies, and community affiliations. This increases the probability that nodes with similar social attributes will meet and forward messages. Therefore, it is feasible to construct opportunistic networks based on social attributes. Such opportunistic networks are called social opportunistic networks. When the node carriers are moving vehicles, these networks can also be called social in-vehicle networks. In social opportunistic network models, users are often required to share their trajectory information to calculate their social attributes and access social opportunistic network services. However, the public disclosure of user trajectory data can lead to privacy breaches. This is because trajectories contain sensitive information such as users' living habits, home addresses, and workplaces. A compromise of the server storing user location information could lead to the disclosure of users' social attributes, which attackers could exploit for identity theft and physical surveillance. Therefore, trajectory privacy protection is crucial to providing good social vehicular network services.

[0003] Given the importance of social attributes in social opportunistic networks, this paper proposes a privacy-preserving routing solution for social in-vehicle networks based on local differential privacy, taking into account the social nature of nodes and the privacy of their social attributes. This solution aims to improve the security and transmission performance of social in-vehicle networks. Compared with other methods, this method is different in the following ways:

[0004] 1. Other community-based opportunistic routing algorithms only consider scenarios where nodes belong to a single community. However, this invention calculates the hotspot locations of nodes based on a hotspot location estimation model, clusters them based on the hotspot locations, and divides the nodes into multiple communities, which is more in line with real-world scenarios.

[0005] 2. Other social routing methods generally fail to consider user privacy protection or fail to guarantee data availability while protecting user privacy. This paper introduces a local differential privacy protection method that adds noise to sensitive data to be accessed, preventing attackers from inferring user privacy while ensuring data availability while maintaining privacy.

[0006] 3. When designing the routing method, we fully considered the situation where nodes belong to multiple communities. Taking into account the difference in the difficulty of node propagation within and between communities, we proposed the concept of vehicle activity to facilitate the propagation of messages between communities. Considering the impact of the number of message copies on the overall performance of opportunistic networks, we proposed a quantitative message copy propagation method to control the total number of copies of a single message. Summary of the invention:

[0007] This paper proposes a privacy protection routing solution for community opportunistic networks, which mainly includes four aspects:

[0008] (1) The clustering algorithm based on AP features calculates the user's social attributes.

[0009] (2) Hotspot location obfuscation algorithm based on local differential privacy.

[0010] (3) Divide vehicle communities based on spectral clustering algorithm and vehicle hotspot locations.

[0011] (4) Message forwarding strategy that integrates the characteristics of multiple user communities.

[0012] The specific contents are as follows:

[0013] (1) The clustering algorithm based on AP features calculates the user's social attributes.

[0014] User mobility patterns are generally considered random, but research shows that user mobility has certain community characteristics and regularity, which generally include the following characteristics. For details, see the attached Figure 1 As shown:

[0015] 1. Users frequently move around a few fixed points, such as work, school, and home.

[0016] 2. Users tend to move along the same path and go back and forth between these points.

[0017] Users' daily commuting points are often relatively fixed, and they will stay near several specific points for a long time. These points that are visited frequently are the users' hot spots and can reflect the users' social attributes.

[0018] The present invention uses the AP (affinity propagation) algorithm to calculate the hotspot location of a single user. The core idea of ​​the algorithm is to regard each data point in the data set as a potential cluster center, called an exemplar. By constructing a network in which each data point is connected to all other data points through edges, a complete similarity matrix is ​​formed. In this network, each edge conveys two pieces of information: responsibility and availability. In this way, the algorithm can calculate the final cluster center of each sample. Common clustering algorithms such as the K-means algorithm must first determine the number of clusters before effective clustering can be performed, or such as the DBSCAN algorithm, it is unable to obtain the center point of the cluster, making it difficult to truly apply to the estimation of user hotspot locations. Compared with other clustering algorithms, the main advantage of the AP algorithm in estimating user hotspot locations is that it can simultaneously calculate the number of clusters k and the center point of each cluster, which is more effective in distance-based clustering.

[0019] The AP algorithm doesn't filter out insufficient or low-density data points during calculations. These points can affect the accuracy and subsequent computational performance of the AP algorithm. Therefore, the algorithm needs to be improved to filter out these points before calculations. Vehicles spend most of their time stationary, which often results in a large number of duplicate nodes in vehicle trajectory records. While these duplicate nodes are useful for calculating vehicle hotspot locations, they can also affect algorithm performance, so they need to be properly processed.

[0020] The user's GPS trajectory dataset is P1(x1,x2,…,x n ), cluster the point set using the improved AP algorithm, divide it into K clusters and obtain the center point of each cluster. The specific description is as follows:

[0021] 1. Use the hash counting algorithm to count duplicate nodes and record them as a dictionary D1(x i ,num i ), controls the number of nodes entering the algorithm, and after filtering out duplicate nodes, the data set is P2(x1,x2,…,x m ).

[0022] 2. Input the cluster distance threshold distance and cluster number threshold MinNs to calculate the distance between points.

[0023] 3. According to the cluster distance threshold distance and cluster number threshold MinNs, the points in the data set that do not meet the requirements are filtered out. The filtered data set is P3, and the corresponding dictionary is D2.

[0024] 4. Based on the filtered dataset P3, calculate the negative Euclidean distance between two points and construct a similarity matrix S(i, j). When i = j, it represents the reference degree of the point. This value is initially set to the median of the similarity matrix. For points with multiple duplicate positions, it is initially multiplied by the num corresponding to the dictionary D2. Initialize the attraction matrix R(i, j) and the attribution matrix A(i, j). Both R and A are initially all zero matrices.

[0025] 5. Enter the maximum comparison value compare, the damping coefficient lam, and the center confirmation threshold Ra.

[0026] 6. Calculate the attraction matrix R(i, j) and the attribution matrix A(i, j) and iterate. The attraction iteration formula and the attribution iteration formula are:

[0027] R t+1 (i,k)=(1-λ)·R t+1 (i,k)+λ·R t (i,k)

[0028]

[0029] A t+1 (i,k)=(1-λ)·A t+1 (i,k)+λ·A t (i,k)

[0030]

[0031] Description: and R t+1 and A t+1 It represents the attraction and belonging degree of the new iteration, and λ represents the damping coefficient, which is used for the convergence of the algorithm.

[0032] 7. Determine whether k meets the requirements and update the hotspot position sequence set Result(x1,x2,…,x m ).

[0033] 8. Repeat steps 6 and 7. If the sequence Result does not change after consecutive compare iterations, the iteration ends and the hotspot location set Result is output. The clustering effect is estimated as shown in the attached figure. Figure 2 shown.

[0034] (2) Hotspot location obfuscation algorithm based on local differential privacy mechanism.

[0035] Hotspot locations reflect users' social attributes and are considered private information. If the central server used to process and allocate communities is untrustworthy, users' hotspot locations will be exposed, and their privacy will be threatened. Excessive data processing will affect the availability of hotspot data and reduce the performance of the vehicle network. This paper uses a local differential privacy method to add noise to hotspot locations, protecting users' privacy while ensuring the availability of hotspot locations. Selecting real nodes in the same cluster as obfuscation targets has two characteristics:

[0036] 1. Real locations do not require availability analysis because real nodes already exist and will not appear in unavailable locations. They are also close to hotspots and have similar social characteristics. The obfuscation system does not need to introduce additional road networks as a reference, which speeds up the obfuscation process.

[0037] 2. Using real data means making trade-offs in security and posing a certain risk of privacy leakage. For location points in the same cluster, some locations have lower density and shorter user stay times, resulting in less privacy leakage. This needs to be considered when performing obfuscation.

[0038] Therefore, choosing the right real location for obfuscation is a challenge. This paper proposes a personalized local differential privacy obfuscation scheme based on point distance and density, maximizing the balance between data security and availability. The details are as follows:

[0039] 1. Input a single hotspot location HP and a dictionary of points D that are in the same cluster as the hotspot location HP (x i ,num i ).

[0040] 2. Calculate the Gaussian kernel function of the distance from each point to the hotspot location, count the number of points within the nearby distance range, and calculate the probability score of each point based on the scoring function Q. The scoring function is as follows:

[0041]

[0042] where num i It is point x i The number of in the dictionary, σ is the hyperparameter of the Gaussian kernel function.

[0043] 3. Use the Laplace mechanism to confuse each probability score. The confusion function is as follows:

[0044]

[0045] where ε q It is the allocated privacy budget used to regulate the privacy strength. The larger the value, the more noise is added and the stronger the privacy protection.

[0046] 4. Calculate the confusion probability based on the obfuscated score, obfuscate the hotspot location to other points in the same cluster based on the probability, and return the obfuscated location. The obfuscation effect is shown in the attached figure. Figure 3 shown.

[0047] (3) Dividing vehicle communities based on spectral clustering algorithm and vehicle hotspot locations

[0048] The routing of the present invention mainly includes two parts. In the first part, users need to upload their own obfuscated hotspot location data to the central server. The central server assigns communities to vehicles according to the hotspot locations of different vehicles. In the second part, users selectively store and forward messages in the self-organizing network according to the community, making full use of the efficient forwarding characteristics of the user community, improving the forwarding success probability, and reducing the forwarding delay.

[0049] The hotspot locations of different users may be close or consistent, and a user's hotspot location often contains more than one. Based on this, the present invention uses a clustering algorithm to divide the user into communities. The spectral clustering algorithm is widely used in classification algorithms. Compared with the traditional K-Means clustering algorithm, the spectral clustering algorithm has a stronger adaptability to data, better calculation effect, smaller amount of calculation, and is not complicated to implement. The spectral clustering algorithm is based on graph theory. Its purpose is to find a method for cutting the graph to ensure that each subgraph is highly similar after cutting, and the similarity between subgraphs is very low. The present invention mainly divides the community based on the distance between the hotspot locations. Therefore, the Gaussian kernel function is selected as the similarity function, the hotspot location of the vehicle is used as the node, and the similarity between points is used as the weight of the edge. The undirected weighted graph G = (V, E) is constructed, where E is the edge set and V is the point set. The algorithm is described as follows:

[0050] 1. Calculate the similarity matrix W. Use the Gaussian kernel function between two points to calculate the similarity and construct the similarity matrix W, where

[0051] The Gaussian kernel function definition formula is:

[0052]

[0053] 2. Calculate the degree matrix D. Only the diagonal of the degree matrix D has valid values, and the vertex V i The degree is the sum of the i-th row of the similarity matrix W, and its definition formula is:

[0054]

[0055] 3. Calculate the Laplace matrix L = DW and normalize the Laplace matrix.

[0056]

[0057] 4. Calculate the smallest k eigenvalues ​​of Q and the corresponding eigenvector f, standardize f, and form an n×k characteristic matrix F.

[0058] 5. Treat each row in F as a sample, for a total of n samples, and cluster the samples using the k-means algorithm.

[0059] 6. Obtain the community division C and assign a community ID to the vehicle.

[0060] (4) A message forwarding strategy that integrates users’ multi-community characteristics and activity levels.

[0061] Because other inventions often assume that users belong to a single community when selecting routing nodes, this significantly reduces the number of next-hop nodes available, inhibiting the likelihood of message forwarding and hindering the success rate of message forwarding. For community-based forwarding routing algorithms, the success rate of message forwarding is significantly improved after reaching the target community. However, improving this success rate is a critical issue.

[0062] Research shows that vehicles are stationary two-thirds of the time during a day, and the probability of encountering other vehicles is significantly reduced when these vehicles are stationary. Furthermore, traffic conditions at the vehicle's location also affect the probability of successful message forwarding. For moving vehicles, there are several scenarios: Vehicle A is traveling in an urban area, where traffic congestion causes the vehicle to stop for extended periods. However, due to heavy traffic, vehicle A is more likely to successfully forward messages to vehicles in the target community. Vehicle B is traveling in a suburban area, where traffic is light and the vehicle travels a greater distance, making it more likely to successfully forward messages between communities. Therefore, the vehicle's activity level is determined by combining the vehicle's travel distance and the number of vehicles it encounters. The greater the activity level, the greater the probability of successful message forwarding between communities. The activity level (liv) is defined as follows.

[0063]

[0064] Where d is the total distance traveled by the vehicle over the past period of time, and n is the number of vehicles encountered by the vehicle over the past period of time. α and β are weight parameters used to adjust the contribution of the total distance and the number of vehicles to liv. d and σ n is the standard deviation of the total distance and number of vehicles.

[0065] When two users meet, they first exchange the community information com (the community set to which the current node belongs), the activity information liv (the centrality of the current node), and the message list mes (summary information of the message carried: message ID, destination node, community ID to which the destination node belongs, and the number of remaining hops N of the node). The com+liv+mes information is used to determine whether to forward the message to the other party. The specific details of the system model are shown in the attached figure. Figure 4 shown.

[0066] In order to limit the number of messages in the space and prevent network space congestion caused by excessive message replication, it is necessary to limit the number of times each message can be forwarded, and use a limited step size method to process and forward messages, that is, add a remaining step size attribute to each message to help forwarding vehicles decide on the message forwarding strategy. Once a message is sent, the number of remaining node hops N of the message is reduced by 1, and the message stored in the vehicle is consistent with the N of the copy message. A vehicle that receives a message with a forwarding count of 0 will only send the message once again, and then the message will be deleted from the vehicle cache. The message will not stop being forwarded until it is transmitted to the target node. This ensures that the number of copies of a message in the network will not grow disorderly, and the total number of copies will be less than N*(N+1) / 2. The specific process is as follows. Figure 5 shown.

[0067] Assume that user A has a message (ID: m, destination user: B, target user community Cb = {C1, C2, C3}, remaining hops: N). When encountering user D:

[0068] If the remaining hop count of the message is greater than 0 and Cb^Cd is not equal to null, a copy of the message is forwarded to user D, and the remaining hop count of the message is reduced by 1.

[0069] If the number of remaining hops of the message is equal to 0 and Cb^Cd is not equal to null, if user D's activity is greater than user B, then a copy of the message is forwarded to user D, and user B deletes the local copy of the message.

[0070] In other cases, the message will not be forwarded. Description of the drawings:

[0071] Figure 1 Move example diagram for users.

[0072] Figure 2 A comparison chart of the clustering effects of the hotspot location estimation model.

[0073] Figure 3 This is a confusing effect diagram.

[0074] Figure 4 System model diagram.

[0075] Figure 5Create a flow chart for the message forwarding strategy. Specific implementation method:

[0076] The present invention is a privacy-preserving routing solution for community vehicle networks based on local differential privacy. The specific steps are as follows:

[0077] Step 1: Determine the hotspot location

[0078] User movement patterns are highly regular; users often stop at a few fixed locations and travel back and forth between them. By analyzing these patterns and mining users' social attributes, we can assist in message routing within the in-vehicle network. In this solution, we use a modified affinity propagation algorithm to build a user hotspot location prediction model to identify user hotspot locations.

[0079] Step 2: Protecting hotspot locations

[0080] Hotspot locations can reflect the social attributes of users and are personal privacy content for users. If the central server used to process and distribute communities is untrustworthy, the user's hotspot location will be exposed and the user's privacy will be threatened. Excessive processing of data will affect the availability of hotspot data and reduce the performance of the vehicle network. The present invention uses a local differential privacy method to add noise to the hotspot location, protecting the user's privacy while ensuring the availability of the hotspot location. The points in the same cluster as the hotspot location in step one are selected as the confusion point set, and each point is scored based on the distance between its nearest neighbors and the distance to the hotspot location, and the point score is numerically noised using the Laplace mechanism. The probability of the point being confused is confirmed based on the point score, and the hotspot location is confused based on the exponential mechanism.

[0081] Step 3: Community Division

[0082] Since each user has one or more hotspot locations, different users may have the same hotspot location, or their hotspot locations may be close to each other. Therefore, a spectral aggregation algorithm is used to cluster the hotspot locations of all users. Users are then divided into communities based on the hotspot clustering results. A user may have more than one hotspot location, so the same user may belong to multiple different communities.

[0083] Step 4: Message delivery

[0084] Messages within the same community will spread easily, while messages between different communities need to consider the centrality of the nodes. Therefore, the routing method of the present invention not only gives priority to nodes with the same community, but also fully considers the impact of the node centrality on the success rate of message delivery. When two nodes meet, if the node exists in the community of the message target vehicle, delivery will be performed. For vehicles with a centrality greater than itself, delivery can be selected even if there is no common community. If there is no node that meets the requirements, it will continue to move and will not forward the message. Taking into account the performance impact of message copies on network space, in order to prevent the disorderly growth of message copies, the routing method determines the number of times the message can be forwarded. When forwarding to a vehicle with the same community, the forwardable flag is cleared to 0, and a copy can only be forwarded once and then deleted. When forwarding to a vehicle without a common community, the number of times the message can be forwarded is reduced by 1.

Claims

1. A privacy-preserving routing solution for social in-vehicle networks based on local differential privacy, which features include: (1) Calculate the user's social attributes based on the clustering algorithm based on AP features; (2) Hotspot location obfuscation algorithm based on exponential mechanism local differential privacy; (3) Divide vehicle communities based on spectral clustering algorithm and vehicle hotspot locations; (4) Message forwarding strategy that integrates the characteristics of multiple user communities; The characteristics of the (1) clustering algorithm based on AP features for calculating the social attributes of users are: using the AP clustering algorithm to solve the identification and number problems of user social hotspots. The AP algorithm can directly identify the K value, that is, the number of user hotspot locations, and can also identify the specific coordinates of the user's hotspot locations, where the particle location is the user's hotspot location; The characteristics of the hotspot location obfuscation algorithm (2) based on local differential privacy are as follows: after the user's hotspot location is calculated by feature (1), each hotspot location needs to be obfuscated to enhance privacy. First, the points in the same cluster as the hotspot location are scored. The scoring is based on the distance of the node from the hotspot location and the node residence time. The higher the score, the greater the probability of being obfuscated. Then, the node score is obfuscated using local differential privacy, and the obfuscation probability is calculated based on the score. The final hotspot location is confirmed based on the probability, thereby ensuring user privacy while ensuring data availability. The feature of (3) dividing vehicle communities based on the spectral clustering algorithm and vehicle hotspot locations is: the hotspot locations that are safe and available to users are obtained through features (1) and (2). Since the hotspot locations of different users are the same or very close, the spectral clustering algorithm is used to cluster all user locations again to divide the communities to which the users belong; The characteristics of the message forwarding strategy (4) that integrates the characteristics of multiple user communities are: The community to which the user belongs is obtained through feature (3), and the community is used as one of the credentials for forwarding messages. When two users meet, they exchange their respective community and activity information. For messages with a remaining hop count greater than 0, if the target user of the message and the encountered user exist in the same community, the message can be forwarded, and the hop count of the message is reduced by 1. For messages with a remaining hop count greater than 0, if the target user of the message and the encountered user exist in the same community and the activity of the encountered vehicle is greater than the current vehicle activity, the message is forwarded and the local copy of the message is deleted.

Images