Federal learning method and system for privacy protection in mobile crowd sensing

Abstract

The invention discloses a federated learning method and system for privacy protection in mobile crowd sensing, and belongs to the field of computer security. The problems that in an existing federation learning method, a server deduces sensitive data of a client side according to gradient parameters, the gradient parameters uploaded by the client side are prone to being intercepted by attackers, user privacy is leaked, and computing resources are wasted in the process that the server waits for a network to delay the client side are solved. According to the method, the parameters of the clients are encrypted by using the semi-homomorphic encryption algorithm based on Paillier, the server side can only obtain the sum of the gradient parameters sent by the clients, the problem that the server deduces the sensitive data of the clients according to the corresponding gradient parameters is avoided, and the sensitive data of the clients can be obtained by setting the fixed aggregation waiting time. Robustness of mobile crowd sensing is enhanced, sensing tasks can still be completed in scenes of user offline, network faults and the like, and waste of computing resources is avoided. The method can be applied to privacy protection in mobile crowd sensing.

Description

Technical Field

[0001] The present invention belongs to the field of computer security, and in particular relates to a federal learning method and system for privacy protection in mobile crowd intelligence perception. Background Art

[0002] With the development of communication technology, mobile crowd intelligence perception, as a way to collect user data on a large scale, has greatly improved the richness and real-time performance of various applications. As mobile crowd intelligence perception develops, the privacy of user data is becoming more and more important. When users enjoy various applications based on mobile crowd intelligence perception, they are also prone to leaking privacy, such as sensitive information such as account balances. If the data is not processed, it may give criminals an opportunity to take advantage of it.

[0003] As a distributed machine learning algorithm, the main idea of ​​federated learning is to ensure that the data of the participants is retained locally, and only the trained model parameters are uploaded to the server for aggregation. This is consistent with the purpose of protecting user data privacy in mobile crowd sensing. Therefore, the use of federated learning in mobile crowd sensing can well protect user privacy data. However, the current federated learning still has certain limitations. Although the server cannot directly obtain user data, its honest and curious characteristics make the server often try to deduce user sensitive data through the gradients or model parameters uploaded by the user while executing the algorithm protocol. The gradient parameters uploaded by the client are easily intercepted by attackers, resulting in user privacy leakage. Moreover, the client that is susceptible to the network or user's subjective influence is prone to quit midway or delay the upload of parameters, and the server is prone to waste computing resources while waiting for the client to upload parameters. Based on the above limitations, it is urgent to propose a federated learning algorithm with privacy protection characteristics and comprehensive performance. Summary of the invention

[0004] The purpose of the present invention is to solve the problems that in the current federated learning method, the server will infer the sensitive data of the client based on the gradient parameters, the gradient parameters uploaded by the client can be easily intercepted by the attacker, resulting in the leakage of user privacy, and the process of the server waiting for the network to delay the client will cause a waste of computing resources. A federated learning method and system for privacy protection in mobile crowd intelligence perception are proposed.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] According to one aspect of the present invention, a privacy-preserving federated learning method in mobile crowdsensing is provided, the method specifically comprising the following steps:

[0007] Step 1: The trusted center generates a key consisting of a public key pk and a private key sk, and publishes the public key pk to all clients and servers, and publishes the private key sk to the server;

[0008] Each client sends an ack to the server to confirm whether to participate in the federated training. The server sends the client set U that confirms to participate in the federated training to the trusted center, and each client in the client set U is recorded as C1, C2, ..., C n , n is the number of clients in U, and the trusted center assigns a random number x to each client in the client set U i , each client uses the random number assigned to it as its own private key;

[0009] Step 2: Each client in the client set U downloads the global model from the server, and the server sends the initial global model parameters to each client in the client set U. And initialize the training round t=0;

[0010] Step 3: Each client in the client set U performs the tth round of training locally, and obtains the gradient parameters of each client respectively.

[0011] Then update the client's gradient parameters according to the communication status between each client and the server in the client set U, and then use the public key pk and private key x i After encrypting the gradient parameters, an encrypted result is obtained; and the encrypted result is sent to the server;

[0012] Step 4: The server aggregates the received encryption results to obtain the aggregated results, decrypts the aggregated results using the private key sk, and decrypts the global model parameters based on the decrypted results. Update and obtain the updated global model parameters

[0013] The updated global model parameters After encryption, the encrypted global model parameters are sent to each client in the client set U;

[0014] Step 5: Each client in the client set U decrypts the received global model parameters and updates the local model parameters according to the decryption result;

[0015] Step 6: Set t=t+1 and return to step 3 until the model converges, that is, the entire federated learning process is completed.

[0016] Furthermore, the trusted center generates keys using the Paillier key generation system;

[0017] The public key pk is (N, g, H), and the private key sk is (λ, μ);

[0018] Where N = p × q, p and q are both prime numbers; λ = lcm(p-1, q-1), lcm is the least common multiple of p-1 and q-1, H is a hash function, {0, 1} * →Z * N , {0, 1} * Represents a binary sequence of any length, Z * N Represents the set of all integers from 1 to N-1 whose greatest common factor with N is 1; randomly selected Represents 1 to N 2 -1 and N 2 The set of all integers whose greatest common factor is 1;

[0019] gcd(L(g λ mod N 2 ), N)=1, gcd represents L(g λ mod N 2 ) and the greatest common divisor of N, g λ mod N 2 Indicates g λ Divide by N 2 The remainder of L(·) is a function; μ=(L(g λ mod N 2 )) -1 mod N.

[0020] Furthermore, the function L(·) is:

[0021]

[0022] Here, x is the function's independent variable.

[0023] Furthermore, the trusted center assigns a random number x to each client in the client set U. i satisfy:

[0024] ∑ i∈U x i =0mod N

[0025] Here, 0 mod N represents the remainder when 0 is divided by N.

[0026] Furthermore, the gradient parameter for:

[0027]

[0028] Where η is the learning rate, is the model parameter obtained by the i-th client after the t-th training round, D iis the local dataset of the ith client, is the loss function, express The gradient of

[0029] The client's gradient parameters are updated according to the communication status between each client and the server in the client set U, and then the public key pk and the private key x are used to i After encrypting the gradient parameters, the encrypted result is obtained; and the encrypted result is sent to the server, specifically:

[0030] Step 1: Update the gradient parameters of each client according to the communication status between each client in the client set U and the server, that is, according to whether the gradient parameters of each client can be successfully received within the aggregation waiting time T of the server;

[0031] If the gradient parameter of the i-th client can be received within the aggregation waiting time T of the server, then execute step 2;

[0032] If the gradient parameters of the i-th client cannot be received within the aggregation waiting time T of the server, the gradient parameters of the i-th client will not be updated. After the i-th client receives the global model parameters, it will continue to participate in the t+1th round of training locally;

[0033] Step 2: Set the maximum number of rounds l that the client is allowed to exit max , the client set corresponding to the gradient parameters that can be received within the server's aggregation waiting time T is recorded as D;

[0034] For the j-th client in set D, if after the t-1th round of training, the server can receive the gradient parameters of the j-th client within the aggregation waiting time T, then the gradient parameters of the j-th client are directly As the updated gradient parameter of the jth client Using public key pk and private key x i Gradient parameter for the jth client After encryption, the encrypted result Send to the server;

[0035] For the j-th client in set D, if after the t-1th round of training, the server does not receive the gradient parameters of the j-th client within the aggregation waiting time T, then execute step 3;

[0036] Step 3: Before the tth round of training, the number of training rounds in which the server has not received the jth client gradient parameter is recorded as l j If l j Greater than l max , then the gradient parameters of the jth client are not processed;

[0037] If l j Less than or equal to l max , then update the gradient parameters of the jth client to obtain the updated gradient parameters

[0038]

[0039] Then update the gradient parameters After encryption, the encrypted result Send to the server.

[0040] Furthermore, the use of public key pk and private key x i Gradient parameter for the jth client Encryption is performed as follows:

[0041]

[0042] in, is the encrypted result of the j-th client gradient parameter, Enc(·) represents encryption, r j represents the random number selected by the jth client, and H(pk) represents the mapping of the binary sequence to the set Z * N Integers in x representing the integer to be mapped j Second power, express Divide by N 2 The remainder of .

[0043] Furthermore, the server aggregates the received encryption results to obtain an aggregated result, specifically:

[0044]

[0045] Where I is the set of clients corresponding to the encryption results received by the server, is the encryption result of the kth client in set I, C t is the aggregation result.

[0046] Furthermore, the aggregation result is decrypted using the private key sk, specifically:

[0047]

[0048] Here, Dec(·) represents decryption.

[0049] Furthermore, the global model parameters are adjusted based on the decryption result. Update, specifically:

[0050]

[0051] Where |I| is the number of clients in set I.

[0052] According to another aspect of the present invention, a privacy-preserving federated learning system in mobile crowd intelligence perception includes a mobile crowd intelligence perception platform, multiple clients, a server and a trusted center; wherein:

[0053] The trusted center is used to initialize the key, send the public key in the key to multiple clients and servers, and send the private key sk to the server;

[0054] The server is used to confirm multiple client sets participating in federated learning, and send the multiple client sets participating in federated learning to the trusted center, and the trusted center is responsible for allocating private keys to each client participating in federated learning;

[0055] The mobile crowd-sensing platform is responsible for issuing sensing tasks to each client participating in federated learning;

[0056] Each client participating in the federated learning downloads the global model from the server according to the perception task, and uses local data to train the downloaded local model, and after encrypting the trained gradient parameters, uploads the encrypted results to the server;

[0057] The server is also used to aggregate the received encryption results, decrypt the aggregated results according to the private key sk, update the local model parameters according to the decrypted results, and then encrypt the updated local model parameters and send them to each client participating in the federated learning.

[0058] The beneficial effects of the present invention are:

[0059] The present invention is based on the security and efficiency of federated learning and semi-homomorphic encryption, which greatly improves the privacy and utility of mobile crowd intelligence perception. The client's parameters are encrypted using a Paillier-based semi-homomorphic encryption algorithm, and the server can only obtain the sum of the gradient parameters sent by each client, avoiding the problem that the server will infer the client's sensitive data based on the corresponding gradient parameters. In addition, by setting a fixed aggregation waiting time, the robustness of mobile crowd intelligence perception can be enhanced. In scenarios such as user offline and network failure, the perception task can still be completed, the model can be converged safely and accurately, and the waste of computing resources can be avoided.

[0060] At the same time, by encrypting the gradient parameters uploaded by the client, it can avoid being intercepted by attackers and causing user privacy leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is an architecture diagram of a privacy-preserving federated learning system in mobile crowd intelligence perception of the present invention. DETAILED DESCRIPTION

[0062] Specific implementation method 1: A privacy-preserving federated learning method in mobile crowd sensing described in this implementation method specifically includes the following steps:

[0063] Step 1: The trusted center generates a key consisting of a public key pk and a private key sk, and publishes the public key pk to all clients and servers, and publishes the private key sk to the server;

[0064] Each client sends an ack to the server to confirm whether to participate in the federated training. The server sends the client set U that confirms to participate in the federated training to the trusted center, and each client in the client set U is recorded as C1, C2, ..., C n , n is the number of clients in U, and the trusted center assigns a random number x to each client in the client set U i , each client uses the random number assigned to it as its own private key;

[0065] Step 2: Each client in the client set U downloads the global model from the server, and the server sends the initial global model parameters to each client in the client set U. And initialize the training round t=0;

[0066] Step 3: Each client in the client set U performs the tth round of training locally. During the training process, stochastic gradient descent and the local dataset D are used. i Update the model parameters and obtain the gradient parameters of each client

[0067] Then update the client's gradient parameters according to the communication status between each client and the server in the client set U, and then use the public key pk and private key x i After encrypting the gradient parameters, an encrypted result is obtained; and the encrypted result is sent to the server;

[0068] Step 4: The server aggregates the received encryption results to obtain the aggregated results, decrypts the aggregated results using the private key sk, and decrypts the global model parameters based on the decrypted results. Update and obtain the updated global model parameters

[0069] The updated global model parameters are encrypted using a symmetric encryption algorithm (AES) and a public key g. After encryption, the encrypted global model parameters are sent to each client in the client set U;

[0070] Step 5: Each client in the client set U uses the key g to decrypt the received global model parameters, and updates the local model parameters according to the decryption result;

[0071] It can prevent adversaries from inferring user privacy from global model parameters;

[0072] Step 6: Set t=t+1 and return to step 3 until the model converges, that is, the entire federated learning process is completed.

[0073] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the trusted center generates keys using the Paillier key generation system;

[0074] The public key pk is (N, g, H), and the private key sk is (λ, μ);

[0075] Wherein, N = p × q, p and q are both prime numbers, and p and q are both prime numbers of 1024 bits or more; λ = lcm(p-1, q-1), lcm is the least common multiple of p-1 and q-1, H is a hash function (specifically, SHA-256 can be selected), {0, 1} * →Z * N , {0, 1} * Represents a binary sequence of any length, Z * N Represents the set of all integers from 1 to N-1 whose greatest common factor with N is 1; randomly selected Represents 1 to N 2 -1 and N 2 The set of all integers whose greatest common factor is 1;

[0076] gcd(L(g λ mod N 2 ), N)=1, gcd represents L(g λ mod N 2 ) and the greatest common divisor of N, g λ mod N 2 Indicates g λ Divide by N 2 The remainder of L(·) is a function; μ=(L(g λ mod N 2 )) -1 mod N.

[0077] The other steps and parameters are the same as those in the first embodiment.

[0078] Specific implementation method three: This implementation method is different from specific implementation method one or two in that the function L(·) is:

[0079]

[0080] Here, x is the function's independent variable.

[0081] The other steps and parameters are the same as those in the first or second embodiment.

[0082] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that the trusted center assigns a random number x to each client in the client set U. i satisfy:

[0083] ∑ i∈U x i =0mod N

[0084] Here, 0 mod N represents the remainder when 0 is divided by N.

[0085] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0086] Specific implementation method 5: This implementation method is different from the specific implementation methods 1 to 4 in that the gradient parameter for:

[0087]

[0088] Where η is the learning rate, is the model parameter obtained by the i-th client after the t-th training round, D i is the local dataset of the ith client, is the loss function, express The gradient of

[0089] The client's gradient parameters are updated according to the communication status between each client and the server in the client set U, and then the public key pk and the private key x are used to i After encrypting the gradient parameters, the encrypted result is obtained; and the encrypted result is sent to the server, specifically:

[0090] Step 1: Update the gradient parameters of each client according to the communication status between each client in the client set U and the server, that is, according to whether the gradient parameters of each client can be successfully received within the aggregation waiting time T of the server;

[0091] If the gradient parameters of the i-th client can be received within the aggregation waiting time T of the server (i.e., the i-th client can successfully send the gradient parameters to the server S after the t-th round of training), then execute step 2;

[0092] If the gradient parameters of the i-th client cannot be received within the server's aggregation waiting time T (i.e., the i-th client cannot successfully send the gradient parameters to the server S after the t-th round of training), the gradient parameters of the i-th client will not be updated. After the i-th client receives the global model parameters, it will continue to participate in the t+1th round of training locally.

[0093] Step 2: Set the maximum number of rounds l that the client is allowed to exit max , the client set corresponding to the gradient parameters that can be received within the server's aggregation waiting time T is recorded as D;

[0094] For the j-th client in set D, if after the t-1th round of training, the server can receive the gradient parameters of the j-th client within the aggregation waiting time T, then the gradient parameters of the j-th client are directly As the updated gradient parameter of the jth client Using public key pk and private key x i Gradient parameter for the jth client After encryption, the encrypted result Send to the server;

[0095] For the j-th client in set D, if after the t-1th round of training, the server does not receive the gradient parameters of the j-th client within the aggregation waiting time T, then execute step 3;

[0096] Step 3: Before the tth round of training, the number of training rounds in which the server has not received the jth client gradient parameter is recorded as l j (If t = 0, then l j is 0); if l j Greater than l max , indicating that the model update of the j-th client has lost the meaning of aggregation, and the gradient parameters of the j-th client are not processed (the server does not use the gradient parameters of the j-th client for aggregation in this round, and the j-th client re-participates in the next round of training after receiving the global parameters sent by the server S);

[0097] If l j Less than or equal to l max , then the gradient parameter of the jth client is updated to obtain the updated gradient parameter Δ

[0098]

[0099]

[0100] Then update the gradient parameters After encryption, the encrypted result Send to the server.

[0101] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0102] If the client fails to upload the gradient parameters to the server in time due to certain factors (such as network delay, client exiting midway, etc.), the server will enter a waiting state. This implementation method can avoid the waste of time and computing resources caused by entering the waiting state.

[0103] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that the public key pk and the private key x are used. i Gradient parameter for the jth client Encryption is performed as follows:

[0104]

[0105] in, is the encrypted result of the j-th client gradient parameter, Enc(·) represents encryption, r j represents the random number selected by the jth client, and H(pk) represents the mapping of the binary sequence to the set Z * N Integers in x representing the integer to be mapped j Second power, express Divide by N 2 The remainder of .

[0106] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0107] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the server aggregates the received encryption results to obtain an aggregated result, specifically:

[0108]

[0109] Where I is the set of clients corresponding to the encryption results received by the server, is the encryption result of the kth client in set I, C t is the aggregation result.

[0110] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0111] The Paillier algorithm is homomorphic, Enc(a)×Enc(b)=Enc(a+b), so the encrypted sum of parameter updates can be obtained by using this implementation.

[0112] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the aggregation result is decrypted using the private key sk, specifically:

[0113]

[0114] Here, Dec(·) represents decryption.

[0115] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0116] Specific implementation method 9: This implementation method is different from any one of the specific implementation methods 1 to 8 in that the global model parameters are adjusted based on the decryption result. Update, specifically:

[0117]

[0118] Where |I| is the number of clients in set I.

[0119] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0120] Specific implementation method ten: Combination Figure 1 This embodiment describes a privacy-preserving federated learning system in mobile crowd intelligence perception, the system comprising a mobile crowd intelligence perception platform, multiple clients, a server and a trusted center; wherein:

[0121] The trusted center is used to initialize the key, send the public key in the key to multiple clients and servers, and send the private key sk to the server;

[0122] The server is used to confirm multiple client sets participating in federated learning, and send the multiple client sets participating in federated learning to the trusted center, and the trusted center is responsible for allocating private keys to each client participating in federated learning;

[0123] The mobile crowd-sensing platform is responsible for issuing sensing tasks to each client participating in federated learning;

[0124] Each client participating in the federated learning downloads the global model from the server according to the perception task, and uses local data to train the downloaded local model, and after encrypting the trained gradient parameters, uploads the encrypted results to the server;

[0125] The server is also used to aggregate the received encryption results, decrypt the aggregated results according to the private key sk, update the local model parameters according to the decrypted results, and then encrypt the updated local model parameters and send them to each client participating in the federated learning.

[0126] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A privacy-preserving federated learning method in mobile crowdsensing, characterized in that: The method specifically comprises the following steps: Step 1: The trusted center generates a key consisting of a public key pk and a private key sk, and publishes the public key pk to all clients and servers, and publishes the private key sk to the server; Each client sends an ack to the server to confirm whether to participate in the federated training. The server sends the client set U that confirms to participate in the federated training to the trusted center, and each client in the client set U is recorded as C1, C2, ..., C n , n is the number of clients in U, and the trusted center assigns a random number x to each client in the client set U i , each client uses the random number assigned to it as its own private key; Step 2: Each client in the client set U downloads the global model from the server, and the server sends the initial global model parameters to each client in the client set U. And initialize the training round t=0; Step 3: Each client in the client set U performs the tth round of training locally, and obtains the gradient parameters of each client respectively. Then update the client's gradient parameters according to the communication status between each client and the server in the client set U, and then use the public key pk and private key x i After encrypting the gradient parameters, an encrypted result is obtained; and the encrypted result is sent to the server; Step 4: The server aggregates the received encryption results to obtain the aggregated results, decrypts the aggregated results using the private key sk, and decrypts the global model parameters based on the decrypted results. Update and obtain the updated global model parameters The updated global model parameters After encryption, the encrypted global model parameters are sent to each client in the client set U; Step 5: Each client in the client set U decrypts the received global model parameters and updates the local model parameters according to the decryption result; Step 6: Set t=t+1 and return to step 3 until the model converges, that is, the entire federated learning process is completed.

2. According to claim 1, the method for privacy-preserving federated learning in mobile crowdsensing is characterized in that: The trusted center generates keys using the Paillier key generation system; The public key pk is (N, g, H), and the private key sk is (λ, μ); Where N = p × q, p and q are both prime numbers; λ = lcm(p-1, q-1), lcm is the least common multiple of p-1 and q-1, H is a hash function, {0, 1} * →Z * N , {0, 1} * Represents a binary sequence of any length, Z * N represents the set of all integers from 1 to N-1 whose greatest common factor with N is 1; randomly select g∈Z * N2 , Z * N2 Represents 1 to N 2 -1 and N 2 The set of all integers whose greatest common factor is 1; gcd(L(g λ mod N 2 ), N) = 1, gcd represents L(g λ mod N 2 ) and the greatest common divisor of N, g λ mod N 2 Indicates g λ Divide by N 2 The remainder of L(·) is a function; μ=(L(g λ mod N 2 )) -1 mod N.

3. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 2, characterized in that: The function L(·) is: Here, x is the function's independent variable.

4. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 3, characterized in that: The trusted center assigns a random number x to each client in the client set U. i satisfy: ∑ i∈U x i =0mod N Here, 0 mod N represents the remainder when 0 is divided by N.

5. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 4, characterized in that: The gradient parameters for: Where η is the learning rate, is the model parameter obtained by the i-th client after the t-th training round, D i is the local dataset of the ith client, is the loss function, express The gradient of The client's gradient parameters are updated according to the communication status between each client and the server in the client set U, and then the public key pk and the private key x are used to i After encrypting the gradient parameters, the encrypted result is obtained; and the encrypted result is sent to the server, specifically: Step 1: Update the gradient parameters of each client according to the communication status between each client in the client set U and the server, that is, according to whether the gradient parameters of each client can be successfully received within the aggregation waiting time T of the server; If the gradient parameter of the i-th client can be received within the aggregation waiting time T of the server, then execute step 2; If the gradient parameters of the i-th client cannot be received within the aggregation waiting time T of the server, the gradient parameters of the i-th client will not be updated. After the i-th client receives the global model parameters, it will continue to participate in the t+1th round of training locally; Step 2: Set the maximum number of rounds l that the client is allowed to exit max , the client set corresponding to the gradient parameters that can be received within the server's aggregation waiting time T is recorded as D; For the j-th client in set D, if after the t-1th round of training, the server can receive the gradient parameters of the j-th client within the aggregation waiting time T, then the gradient parameters of the j-th client are directly As the updated gradient parameter of the jth client Using public key pk and private key x i Gradient parameter for the jth client After encryption, the encrypted result Send to the server; For the j-th client in set D, if after the t-1th round of training, the server does not receive the gradient parameters of the j-th client within the aggregation waiting time T, then execute step 3; Step 3: Before the tth round of training, the number of training rounds in which the server has not received the jth client gradient parameter is recorded as l j If l j Greater than l max , then the gradient parameters of the jth client are not processed; If l j Less than or equal to l max , then update the gradient parameters of the jth client to obtain the updated gradient parameters Then update the gradient parameters After encryption, the encrypted result Send to the server.

6. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 5, characterized in that: The use of public key pk and private key x i Gradient parameter for the jth client Encryption is performed as follows: in, is the encrypted result of the j-th client gradient parameter, Enc(·) represents encryption, r j represents the random number selected by the jth client, and H(pk) represents the mapping of the binary sequence to the set Z * N Integers in x representing the integer to be mapped j Second power, express Divide by N 2 The remainder of .

7. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 6, characterized in that: The server aggregates the received encryption results to obtain an aggregated result, specifically: Where I is the set of clients corresponding to the encryption results received by the server, is the encryption result of the kth client in set I, C t is the aggregation result.

8. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 7, characterized in that: The aggregation result is decrypted using the private key sk, specifically: Here, Dec(·) represents decryption.

9. The method for privacy-preserving federated learning in mobile crowdsensing according to claim 8, characterized in that: The global model parameters are adjusted based on the decryption result. Update, specifically: Where |I| is the number of clients in set I.

10. A privacy-preserving federated learning system in mobile crowd intelligence perception, characterized in that: The system includes a mobile crowd intelligence perception platform, multiple clients, a server and a trusted center; wherein: The trusted center is used to initialize the key, send the public key in the key to multiple clients and servers, and send the private key sk to the server; The server is used to confirm multiple client sets participating in federated learning, and send the multiple client sets participating in federated learning to the trusted center, and the trusted center is responsible for allocating private keys to each client participating in federated learning; The mobile crowd-sensing platform is responsible for issuing sensing tasks to each client participating in federated learning; Each client participating in the federated learning downloads the global model from the server according to the perception task, and uses local data to train the downloaded local model, and after encrypting the trained gradient parameters, uploads the encrypted results to the server; The server is also used to aggregate the received encryption results, decrypt the aggregated results according to the private key sk, update the local model parameters according to the decrypted results, and then encrypt the updated local model parameters and send them to each client participating in the federated learning.

Images