Privacy protection federated aggregation method based on multi-client inner product function proxy re-encryption
Through the multi-client internal product function proxy re-encryption method, the problem of high computational overhead of central server obtaining global model parameters and realizing the secure aggregation and privacy protection of model parameters, which is suitable for medical data training with strict privacy protection .
Patent Information
- Application Number
- CN202411283138.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2025-01-03
AI Technical Summary
The existing federated learning framework based on multi-input function encryption has the presence of a central server that can combine different rounds of ciphertexts and decryption keys to reveal the security risks of valuable information in the original data, and sharing session keys among participants causes the central server to be able to Obtaining multiple sets of parameters leads to problems with original data leakage and high computational overhead.
Using multi-client internal product function proxy re-encryption method, the decryption key and re-encryption key are generated through the key generation center. The client trains and encrypts the model parameters locally, and the central server re-encrypts and aggregates and then transmits them to ensure semi-honest The server cannot obtain plain text information, reducing client resource overhead.
It effectively protects the privacy of model parameters, reduces the risk of central server obtaining global model parameters, and reduces the computing load of clients. It is suitable for medical model training scenarios with strict privacy protection.
Smart Images
Figure CN119249444A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of privacy protection technology, and in particular to a privacy protection federated aggregation method based on multi-client inner product function proxy re-encryption. Background Art
[0002] In modern medical activities, artificial intelligence represented by machine learning model training and reasoning plays an increasingly important role. In the process of model training, it is often necessary to introduce multi-dimensional medical data. These data may come from multiple users or multiple medical institutions. In the traditional model training process, it is necessary to first collect all the data involved in the training and then conduct centralized training. In the data collection process, due to the cross-domain transmission of data, there is a greater risk of user privacy leakage.
[0003] Federated learning allows different participants to collaborate on model training without transmitting original data, thereby protecting user data privacy. However, federated learning faces the threat of central servers recovering information related to participants' original data through intermediate parameters. To solve this privacy protection problem, there are many ways to aggregate federated learning parameters, such as federated learning based on secure multi-party computation (MPC), in which participants share intermediate parameters using secret sharing, and the central server aggregates parameters by running the MPC protocol with participants, and then restores the global parameters through secret sharing reconstruction, but this introduces more communication rounds and significant overhead. In federated learning based on homomorphic encryption (HE), participants share keys and encrypt intermediate parameters. Although it allows calculations on encrypted data so that the central server can aggregate parameters using homomorphic operations and broadcast the results, the complexity of homomorphic encryption usually brings a lot of computational overhead. In federated learning based on differential privacy (DP), participants add noise to the intermediate parameters and upload them to the central server. The server aggregates random parameters and broadcasts the results, but it is easy to cause model accuracy loss due to a large amount of noise. In contrast, federated learning based on function encryption (FE) better balances security and performance and has stronger practicality. Multi-input function encryption (MIFE) is a promising technology applied to FE-based federated learning, which enables multiple participants to encrypt plaintext. Participants with decryption keys can evaluate functions and the input comes from multiple participants. In federated learning, MIFE is generally used to protect local model parameters from being leaked by the central server, but it cannot prevent the server from obtaining global model parameters. Therefore, the existing federated learning framework based on MIFE has multiple security risks. For example, the central server can combine ciphertext and decryption keys of different rounds to reveal valuable information of the original data, and the session key is shared among participants and used multiple times, resulting in the central server obtaining multiple sets of parameters corresponding to the same session key, thereby obtaining plaintext information by combining these parameters.
[0004] Therefore, it is necessary to provide a privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption to solve the above technical problems. Summary of the invention
[0005] The present invention provides a privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption, which solves the problem that the central server cannot be prevented from obtaining global model parameters, which may lead to the leakage of valuable information in the original data and the large amount of computing overhead usually caused by the complexity of homomorphic encryption, as well as the session key being shared among participants and used multiple times, so that the central server can obtain plaintext information by combining multiple sets of parameters.
[0006] In order to solve the above technical problems, the present invention provides a privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption, which includes the following four steps: initialization phase, local training phase, federated aggregation phase and parameter update phase, which are as follows:
[0007] Initialization phase: Step 110, the key generation center (KGC) inputs the security parameter λ and calls the group generation algorithm, and then selects the hash function Zero i =s i , the public parameters and the master key can be expressed as and msk=((ek i ) i ), KGC broadcasts the public parameter PP and stores the master key msk in a secure area;
[0008] Step 111, the key generation center (KGC) inputs the master key msk=(s i ) i and a vector And returns the decryption key For each client, KGC calculation And let the decryption key be Finally, KGC sends the decryption key to each client;
[0009] Step 112, randomly select a client, the client selects a random number For i∈[m], client j∈[n] computes And send rk to the central server;
[0010] Local training stage: Step 201, each client performs model training locally using local data to obtain trained local model parameters;
[0011] Step 202: for each dimensional model parameter x obtained in step 201 i , client computing Then calculate The ciphertext can be expressed as
[0012] Step 203, each client uploads the parameter ciphertext ct obtained after encryption in step 202 i to the central server;
[0013] Federation aggregation phase: Step 204, the central server collects the model parameter ciphertext sent by each client. For i∈[m], the central server uses the re-encryption key rk to encrypt the original ciphertext ct i Perform re-encryption and calculate ct′;
[0014] Step 205, each client downloads the aggregated model parameter ciphertext ct′ from the central server to the local computer;
[0015] Parameter update phase: Step 206: Each client uses its own decryption key to decrypt the obtained ciphertext, that is, the client with dk = s′ calculates And get the inner product result
[0016] Step 207, each client updates the local model parameters using the inner product result obtained in step 206, and uses the updated parameters to perform the next round of iteration;
[0017] Each client and the central server repeats step 201 to step 207 until the model converges.
[0018] Preferably, the group generation algorithm in step 110 is: Hash functions for You can map any string to Elements of the group.
[0019] Preferably, n in step 202 j is the vector dimension of client j.
[0020] Preferably, m in step 204 is the vector length, and the method for calculating ct′ is:
[0021] Preferably, the central server includes a main body, a plurality of plug interfaces are provided on the front side of the main body, a mounting plate is fixedly installed on the front side of the main body, a mounting groove is provided on the front side of the mounting plate, a sliding groove is provided on the front side of the mounting frame, an abutment piece is slidably connected to the inner side surface of the sliding groove, and a snap-in groove is provided on the top and bottom of the mounting frame.
[0022] Preferably, the top and bottom of the mounting plate are slidably connected to baffles, a sliding cavity is opened inside the baffles, a spring is provided on the inner side of the sliding cavity, a sliding rod is slidably connected inside the baffles, and the end face of the sliding rod is provided on the inner side of the sliding cavity.
[0023] Preferably, a clamping block is fixedly mounted on the outer surface of the baffle, and the clamping block is arranged on one side of the clamping groove.
[0024] Preferably, a through groove is formed at one end of the baffle, and a sponge pad is fixedly mounted on the inner side of the through groove.
[0025] Preferably, the interior of the slide groove and the inner side surface of the clamping groove are connected.
[0026] Preferably, a rotating seat is fixedly mounted on the top and bottom of the front side of the mounting plate, the inner side of the rotating seat is rotatably connected to a rotating block, a protective cover is fixedly mounted on one side of the rotating block, and the protective cover is arranged on the outer side of the abutment member.
[0027] Compared with the related art, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention has the following beneficial effects:
[0028] The present invention provides a privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption. By calculating the parameter aggregation stage where the load is mainly on the central server, the computing power advantage of the central server can be fully utilized to reduce the resource overhead of the client.
[0029] In the process of parameter transmission, the inner product function is used to encrypt the model parameters. At the same time, the idea of proxy re-encryption is used to ensure that the semi-honest central server cannot access the plaintext information of any parameters, thereby ensuring the security of relevant parameters and further ensuring the privacy of local data.
[0030] During the model training process, each participant (client) can use each party's data for model training without transmitting their own local sensitive data. This allows for more extensive use of multi-party data to train more accurate models while ensuring data privacy. This is suitable for medical model training scenarios with stricter privacy protection requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 A schematic diagram of the process flow of the initialization phase of the medical data-oriented federated learning method provided by the present invention;
[0032] Figure 2 A schematic flow chart of the local training phase, the federated aggregation phase and the parameter updating phase of the federated learning method for medical data provided by the present invention;
[0033] Figure 3 A schematic diagram of a second implementation structure of the initialization phase of the medical data-oriented federated learning method provided by the present invention;
[0034] Figure 4 for Figure 3 An enlarged schematic diagram of part A is shown;
[0035] Figure 5 for Figure 4 The schematic diagram of the side cross-sectional structure of the baffle shown;
[0036] Figure 6 A schematic diagram of a third implementation structure of the initialization phase of the medical data-oriented federated learning method provided by the present invention;
[0037] Numbers in the figure: 1, main body, 11, plug interface, 2, mounting plate, 21, mounting groove, 22, clamping groove, 23, slide groove, 24, abutment, 3, baffle, 31, clamping block, 32, through groove, 33, sponge pad, 34, slide cavity, 35, spring, 36, limit plate, 37, slide rod, 4, rotating seat, 41, rotating block, 43, protective cover. DETAILED DESCRIPTION
[0038] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.
[0039] First embodiment
[0040] Please refer to Figure 1 , Figure 2 ,in, Figure 1 A schematic diagram of the process flow of the initialization phase of the medical data-oriented federated learning method provided by the present invention; Figure 2 The flowchart of the local training phase, federated aggregation phase and parameter update phase of the federated learning method for medical data provided by the present invention is as follows. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption includes the following four steps: initialization phase, local training phase, federated aggregation phase and parameter update phase, which are as follows:
[0041] Initialization phase: Step 110, the key generation center (KGC) inputs the security parameter λ and calls the group generation algorithm, and then selects the hash function Zero i =s i , the public parameters and the master key can be expressed as and msk=((ek i ) i ), KGC broadcasts the public parameter pp and stores the master key msk in a secure area;
[0042] Step 111, the key generation center (KGC) inputs the master key msk=(s i ) i and a vector And returns the decryption key For each client, KGC calculation And let the decryption key be Finally, KGC sends the decryption key to each client;
[0043] Step 112, randomly select a client, the client selects a random number For i∈[m], client j∈[n] computes And send rk to the central server;
[0044] Local training stage: Step 201, each client performs model training locally using local data to obtain trained local model parameters;
[0045] Step 202: for each dimensional model parameter x obtained in step 201 i , client computing Then calculate The ciphertext can be expressed as
[0046] Step 203, each client uploads the parameter ciphertext ct obtained after encryption in step 202 i to the central server;
[0047] Federation aggregation phase: Step 204, the central server collects the model parameter ciphertext sent by each client. For i∈[m], the central server uses the re-encryption key rk to encrypt the original ciphertext ct i Perform re-encryption and calculate c t ′;
[0048] Step 205, each client downloads the aggregated model parameter ciphertext ct′ from the central server to the local computer;
[0049] Parameter update phase: Step 206: Each client uses its own decryption key to decrypt the obtained ciphertext, that is, the client with dk = s′ calculates And get the inner product result
[0050] Step 207, each client updates the local model parameters using the inner product result obtained in step 206, and uses the updated parameters to perform the next round of iteration;
[0051] Each client and the central server repeats step 201 to step 207 until the model converges.
[0052] The group generation algorithm in step 110 is: Hash functions for You can map any string to Elements of the group.
[0053] The n in step 202 j is the vector dimension of client j.
[0054] In step 204, m is the vector length, and ct′ is calculated as follows:
[0055] Compared with the related art, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention has the following beneficial effects:
[0056] By calculating the parameter aggregation phase where the load is mainly on the central server, the computing power advantage of the central server can be fully utilized to reduce the resource overhead of the client;
[0057] In the process of parameter transmission, the inner product function is used to encrypt the model parameters. At the same time, the idea of proxy re-encryption is used to ensure that the semi-honest central server cannot access the plaintext information of any parameters, thereby ensuring the security of relevant parameters and further ensuring the privacy of local data.
[0058] During the model training process, each participant (client) can use each party's data for model training without transmitting their own local sensitive data. This allows for more extensive use of multi-party data to train more accurate models while ensuring data privacy. This is suitable for medical model training scenarios with stricter privacy protection requirements.
[0059] Second embodiment
[0060] Please refer to Figure 3 , Figure 4 and Figure 5 Based on the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the first embodiment of the present application, the second embodiment of the present application proposes another privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption. The second embodiment is only a preferred method of the first embodiment, and the implementation of the second embodiment will not affect the independent implementation of the first embodiment.
[0061] Specifically, the difference of the privacy-preserving federal aggregation method based on multi-client inner product function proxy re-encryption provided by the second embodiment of the present application is that, based on the privacy-preserving federal aggregation method based on multi-client inner product function proxy re-encryption, the central server includes a main body 1, a plurality of plug interfaces 11 are provided on the front of the main body 1, a mounting plate 2 is fixedly installed on the front of the main body 1, a mounting groove 21 is provided on the front of the mounting plate 2, a slide groove 23 is provided on the front of the mounting frame 2, an abutment member 24 is slidably connected to the inner side surface of the slide groove 23, and a snap-in groove 22 is provided on the top and bottom of the mounting frame 2.
[0062] The top and bottom of the mounting plate 2 are both slidably connected with baffle plates 3, a sliding cavity 34 is provided inside the baffle plates 3, a spring 35 is provided on the inner side of the sliding cavity 34, a sliding rod 37 is slidably connected inside the baffle plates 3, and the end face of the sliding rod 37 is provided on the inner side of the sliding cavity 34.
[0063] A clamping block 31 is fixedly mounted on the outer surface of the baffle 3 , and the clamping block 31 is disposed at one side of the clamping slot 22 .
[0064] A through slot 32 is formed at one end of the baffle plate 3 , and a sponge pad 33 is fixedly mounted on the inner side of the through slot 32 .
[0065] The interior of the slide groove 23 is connected to the inner side surface of the clamping groove 22 .
[0066] There are multiple plug-in ports 11, and baffles 3 are arranged on the top and bottom of each plug-in port 11, and a clamping groove 22 is provided on the top of the mounting plate 2 to cooperate with the clamping block 31, so as to limit the baffle 3;
[0067] The abutment member 24 slides on the inner side of the slide groove 23, so that when in use, one end of the abutment member 24 can push the clamping block 31;
[0068] The through grooves 32 opened on the end faces of the upper and lower baffles 3 can form a circle and cooperate with the sponge pad 33 to wrap the transmission line. At this time, since the size of the connector is larger than the size of the through groove 32, the interface can be blocked to prevent the interface from being pulled out from the inner side of the plug port 11.
[0069] The working principle of the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention is as follows:
[0070] When in use, the user inserts the transmission line interface connected to the body 1 into the inner side of the plug interface 11, and then the user can push the baffles 3 arranged above and below the plug interface 11 toward the middle, so that the slide rod 37 drives the limit plate 36 to slide on the inner side of the slide cavity 34, and then squeezes the spring 35, so that the transmission line is clamped on the inner side of the through groove 32, so that the sponge pad 33 wraps the transmission line, and at this time the clamping block 31 is clamped into the inner side of the clamping groove 22, so that the baffle 3 can be limited;
[0071] When the transmission line needs to be disassembled, the user can press the abutment 24 to make the abutment 24 slide on the inner side of the slide groove 23, and then make one end of the abutment 24 push the clamping block 31, so that the clamping block 31 is deformed and moves out from the inner side of the clamping groove 22. After that, the elasticity of the spring 35 can push the two baffles 3 upward and downward respectively, so as to expose the plug interface 11. At this time, the user can disassemble the transmission line from the top of the plug interface 11.
[0072] Compared with the related art, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention has the following beneficial effects:
[0073] By means of the cooperation of structures such as the mounting plate 2, the mounting groove 21, the snap-in groove 22, the baffle 3, the snap-in block 31, the through groove 32, the sponge pad 33, the sliding cavity 34, the spring 35 and the sliding rod 37, the two baffles 3 can clamp the transmission line when in use, so that the transmission line interface abuts against the inner side surface of the plug-in interface 11, thereby avoiding pulling the transmission line and causing the transmission line interface to fall off from the inner side surface of the plug-in interface 11, thereby increasing the stability of the connection between the transmission line and the body 1.
[0074] Third embodiment
[0075] Please refer to Figure 6 Based on the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the first embodiment of the present application, the third embodiment of the present application proposes another privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption. The third embodiment is only a preferred method of the first embodiment, and the implementation of the third embodiment will not affect the independent implementation of the first embodiment.
[0076] Specifically, the difference of the privacy-preserving federal aggregation method based on multi-client inner product function proxy re-encryption provided in the third embodiment of the present application is that, based on the privacy-preserving federal aggregation method based on multi-client inner product function proxy re-encryption, a rotating seat 4 is fixedly installed on the top and bottom of the front side of the mounting plate 2, and the inner side surface of the rotating seat 4 is rotatably connected to a rotating block 41, and a protective cover 43 is fixedly installed on one side of the rotating block 41, and the protective cover 43 is arranged on the outer side surface of the abutment 24.
[0077] The working principle of the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention is as follows:
[0078] When in use, the protective cover 43 is rotated, so that the protective cover 43 drives the rotating block 41 to rotate inside the rotating seat 4, so that the protective cover 43 covers the outer side of the abutment 24, thereby protecting the abutment 24.
[0079] Compared with the related art, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption provided by the present invention has the following beneficial effects:
[0080] By cooperating with each other, the rotating seat 4, the rotating block 41 and the protective cover 43, when in use, the protective cover 43 is covered on the outer side of the abutment 24, thereby avoiding accidental contact with the abutment 24, causing the clamping block 31 to detach from the inner side of the clamping groove 22, and increasing the stability of the device when in use.
[0081] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption, characterized in that: It includes the following four steps: initialization phase, local training phase, federated aggregation phase and parameter update phase, as follows: Initialization phase: Step 110, the key generation center (KGC) inputs the security parameter λ and calls the group generation algorithm, and then selects the hash function Zero i =s i , the public parameters and the master key can be expressed as and msk=((ek i ) i ), KGC broadcasts the public parameter pp and stores the master key msk in a secure area; Step 111, the key generation center (KGC) inputs the master key msk=(s i ) i and a vector And returns the decryption key For each client, KGC calculation And let the decryption key be Finally, KGC sends the decryption key to each client; Step 112, randomly select a client, the client selects a random number For i∈[m], client j∈[n] computes And send rk to the central server; Local training stage: Step 201, each client performs model training locally using local data to obtain trained local model parameters; Step 202: for each dimensional model parameter x obtained in step 201 i , client computing Then calculate The ciphertext can be expressed as Step 203, each client uploads the parameter ciphertext ct obtained after encryption in step 202 i to the central server; Federation aggregation phase: Step 204, the central server collects the model parameter ciphertext sent by each client. For i∈[m], the central server uses the re-encryption key rk to encrypt the original ciphertext ct i Perform re-encryption and calculate ct'; Step 205, each client downloads the aggregated model parameter ciphertext ct' from the central server to the local computer; Parameter update phase: Step 206: Each client uses its own decryption key to decrypt the obtained ciphertext, that is, the client with dk = s' calculates And get the inner product result Step 207, each client updates the local model parameters using the inner product result obtained in step 206, and uses the updated parameters to perform the next round of iteration; Each client and the central server repeats step 201 to step 207 until the model converges.
2. According to claim 1, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption is characterized in that: The group generation algorithm in step 110 is: Hash Functions for You can map any string to Elements of the group.
3. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 1 is characterized in that: The n in step 202 j is the vector dimension of client j.
4. According to claim 1, the privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption is characterized in that: In step 204, m is the vector length, and the method for calculating ct' is:
5. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 1 is characterized in that: The central server includes a main body, a plurality of plug interfaces are provided on the front of the main body, a mounting plate is fixedly installed on the front of the main body, a mounting groove is provided on the front of the mounting plate, a slide groove is provided on the front of the mounting frame, an abutment piece is slidably connected to the inner side surface of the slide groove, and a snap-in groove is provided on the top and bottom of the mounting frame.
6. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 5 is characterized in that: The top and bottom of the mounting plate are slidably connected with baffles, a sliding cavity is opened inside the baffle, a spring is arranged on the inner side of the sliding cavity, a sliding rod is slidably connected inside the baffle, and the end face of the sliding rod is arranged on the inner side of the sliding cavity.
7. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 6 is characterized in that: A clamping block is fixedly mounted on the outer surface of the baffle, and the clamping block is arranged on one side of the clamping groove.
8. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 7 is characterized in that: A through groove is formed at one end of the baffle, and a sponge pad is fixedly mounted on the inner side of the through groove.
9. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 8 is characterized in that: The interior of the slide groove and the inner side surface of the clamping groove are connected.
10. The privacy-preserving federated aggregation method based on multi-client inner product function proxy re-encryption according to claim 9 is characterized in that: A rotating seat is fixedly mounted on the top and bottom of the front side of the mounting plate, a rotating block is rotatably connected to the inner side of the rotating seat, a protective cover is fixedly mounted on one side of the rotating block, and the protective cover is arranged on the outer side of the abutment member.