Lightweight privacy protection target classification method based on additive secret sharing

Abstract

The invention discloses a lightweight privacy protection target classification method based on additive secret sharing, and the method comprises the steps: encrypting data through an encryption scheme based on Henon and Sine chaotic mapping design, uploading the encrypted data to an edge server, decrypting the data through the edge server, and training a neural network through an efficient and secure calculation protocol based on additive secret sharing design, and a classification framework for protecting data privacy is realized. According to security requirements of a convolutional layer, a pooling layer, an activation layer and a full connection layer in a neural network, an efficient security computing protocol is constructed to realize secure neural network computing. The security computing protocol of the convolutional layer, the pooling layer, the activation layer and the full connection layer in the neural network is fully improved, the communication cost and the computing cost are reduced, and the transmission efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of computing technology, and in particular to an encryption method. Background Art

[0002] Object classification tasks are widely used in scenarios such as finance, medical care, and autonomous driving. With the widespread application of cloud computing, more and more users choose to outsource massive image data and computationally intensive visualization tasks to cloud service centers to improve data processing efficiency. However, since the uploaded data contains a large amount of sensitive information, once leaked, it will pose a serious threat to personal privacy. Therefore, a privacy protection method is urgently needed to ensure security and privacy during data transmission and processing. Existing data privacy protection methods, whether symmetric or asymmetric encryption, will reduce data availability and recognition accuracy, and increase huge computing costs. Homomorphic encryption calculations usually require long data processing delays and are not suitable for application systems that require low latency and high performance.

[0003] Secret sharing technology does not rely on the construction of computationally intensive cryptographic primitives, so it can greatly reduce key management and communication overhead and greatly improve the overall performance of the system. Secret sharing splits data into multiple shares to ensure that the complete data can only be restored when a certain number of participants cooperate, thereby effectively preventing attacks from inside and outside the system and providing strong protection for data security and privacy. Edge computing can transfer heavy computing tasks from mobile devices to edge servers, greatly improving computing efficiency and system response speed to achieve efficient system design, and providing strong support for the construction of a privacy data sharing framework. In recent years, secret sharing and edge computing have attracted much attention, and their many advantages just meet the needs of privacy protection in multiple application scenarios.

[0004] With the accelerated development of wireless communication Internet technology, data security issues have become a focus of industry attention. In June 2021, the promulgation of the "Data Security Law of the People's Republic of China" actively responded to the key issues of data competition and protection at home and abroad, and provided guidance for corporate data management compliance and further data asset governance and development. However, despite the legal protection, how to efficiently solve data security issues is still a major challenge facing the industry.

[0005] To this end, we constructed a lightweight privacy-preserving target classification method based on additive secret sharing, edge computing and chaotic encryption. First, the data provider splits the original data into two independent shares. The third-party lightweight server generates an encryption key based on the chaotic map to encrypt the two shares and sends the encrypted data to the two designated edge servers respectively; then, the edge server performs a decryption operation after receiving the encrypted shares, and implements privacy-preserving neural network training through the designed security function; finally, the edge server sends the result to the data receiver, and the data receiver only needs to add the two results received from the edge server to get the classification result. In this process, the privacy of the original data is effectively protected, and the sharing and utilization of data are realized, thereby realizing privacy-preserving data sharing.

[0006] Regarding the privacy protection target classification method, Hesamifard et al. proposed a solution for training deep neural networks based on encryption technology based on homomorphic encryption. The solution designed a method of using low-order polynomial approximation to replace the original activation function in convolutional neural networks to train convolutional neural networks, thereby ensuring that convolutional neural networks are implemented on the basis of protecting privacy-sensitive original data. Xiong et al. proposed an edge-assisted privacy-preserving original data sharing framework based on additive secret sharing technology. The solution encrypts the original data into two ciphertexts and uses security functions to implement privacy-preserving convolutional neural networks, thereby ensuring data sharing while avoiding the problem of personal information leakage.

[0007] The shortcomings of the prior art are mainly manifested in the following aspects:

[0008] (1) The computational cost and processing delay brought by existing solutions make it difficult to apply them in data sharing application scenarios that require low latency and high timeliness.

[0009] (2) Existing solutions cannot deal with potential adversaries that simultaneously destroy the two transmission channels through which the data sender uploads data. Summary of the invention

[0010] The method proposed by the present invention solves the above problems, which is specifically manifested in:

[0011] (1) Combined with the chaotic encryption scheme, the two data split by the sender are encrypted and uploaded to the edge server, which can effectively prevent both transmission channels from being destroyed by attackers.

[0012] (2) Combined with the Beaver triplet, a lightweight, highly adaptive and efficient secure computing protocol is designed, which effectively reduces communication and computing costs and improves transmission efficiency.

[0013] The technical solution of the present invention:

[0014] A lightweight privacy-preserving target classification method based on additive secret sharing. The encryption scheme designed based on Henon and Sine chaotic mapping is used to encrypt the data and upload it to the edge server. After the edge server decrypts the data, the neural network is trained through an efficient and secure computing protocol designed based on additive secret sharing to implement a classification framework that protects data privacy. Specifically, the following steps are included: Step 1: Initial image generation. The data sender D1 uses a random generator to generate a random image I1, and subtracts I1 from the original image I to obtain I2;

[0015] Step 2: Key generation. Combining Henon mapping and Sine mapping, a new efficient three-dimensional chaotic mapping is constructed as shown in formula (a):

[0016]

[0017] Assume that the image size is (ψ1, ψ2, ψ3). In order to make the constructed chaotic map in a chaotic state, set the value of the control parameter tuple (a, b, c), input (X0, Y0, Z0) to iterate the chaotic map in formula (a) ψ1ψ2ψ3 times, and obtain three chaotic sequences X, Y and Z with lengths of ψ1ψ2ψ3. In order to make the chaotic sequence we obtain periodic, use the sine function to process the above sequence to obtain a chaotic sequence with a longer period in the interval [0,1]. In order to make the random characteristics more ideal and confuse the image pixel values, the sequence values ​​are further confused and the transformation is performed. and Among them, Yanzhong 10 16 To extend the period, round(·) is the operation to the nearest integer;

[0018] Step 3: Image encryption

[0019] We first stretch the images I1 and I2 sent to edge servers S1 and S2 into stream sequences of length ψ1·ψ2·ψ3, and then perform two rounds of encryption in sequence, as shown in equations (b) and (c):

[0020]

[0021] P i (i=1,2) is the encryption sequence after the first round of encryption, B i (i=1,2) is the encryption sequence after the second round of encryption;

[0022] Step 4: Image Decryption

[0023] After edge servers S1 and S2 receive the encrypted images E1 and E2, they first need to perform a reverse encryption operation, where the first round of decryption is the reverse operation of the second round of encryption, and the second round of decryption is the reverse operation of the first round of encryption. The third-party server T provides decryption keys δ1 and δ2. After decryption, images I1 and I2 are obtained. The decryption operation is shown in equations (d) and (e).

[0024]

[0025]

[0026] The linear computing protocols of the secure computing protocol include: secure addition protocol, secure subtraction protocol, and secure scalar multiplication protocol.

[0027] The nonlinear computing protocols of the secure computing protocol include: secure XOR protocol, secure conversion protocol STAX, secure conversion protocol STXA, secure comparison protocol, and secure activation protocol.

[0028] Beneficial effects of the present invention:

[0029] 1. Efficient secure computing protocol. Aiming at the security requirements of convolutional layers, pooling layers, activation layers, and fully connected layers in neural networks, an efficient secure computing protocol is built to achieve secure neural network computing. The secure computing protocol of convolutional layers, pooling layers, activation layers, and fully connected layers in neural networks is fully improved, which not only reduces communication and computing costs, but also improves transmission efficiency.

[0030] 2. High-security chaotic encryption scheme. The chaotic encryption scheme based on Henon and Sine chaotic mapping has the characteristics of strong noise resistance, good chaotic performance and high security, and can effectively prevent the two transmission channels from being damaged by attackers. It can not only effectively prevent potential attackers from damaging the data transmission channel, but also reduce the computational overhead.

[0031] 3. Lightweight object classification framework. Additive secret sharing is used to transfer computationally intensive work to edge servers, thereby significantly reducing users' computing and resource overhead. Transferring computing tasks to edge servers is more suitable for image encryption and transmission in scenarios that require low latency and high performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 1 is a system model diagram according to the present invention. DETAILED DESCRIPTION

[0033] Embodiment 1:

[0034] The invention encrypts data based on an encryption scheme designed by Henon and Sine chaotic mapping and uploads it to an edge server. After the edge server decrypts the data, a neural network is trained through an efficient and secure computing protocol designed based on additive secret sharing to realize a classification framework for protecting data privacy.

[0035] (1) Implementation steps

[0036] Assume that the data sender is D1, the data receiver is D2, the lightweight third-party server is T, the two edge servers are S1 and S2, and I is the original image. The third-party server T randomly generates a random image I1, calculates I2=I-I1, and uses the chaotic encryption scheme to generate encryption keys δ1 and δ2 to encrypt I1 and I2 respectively, and uploads the encrypted images E1 and E2 to the edge servers S1 and S2. The server uses the designed secure computing protocol to extract features and can realize the secure convolution layer, secure fully connected layer, secure activation layer and secure pooling layer, and finally obtains two classification results C1 and C2. The data receiver D2 adds the two classification results to get C=C1+C2, which is the final classification result.

[0037] (2) Encryption Scheme

[0038] The encryption algorithm of the present invention is based on the Henon and Sine chaotic mapping design, and the main steps are as follows:

[0039] Step 1: Initial image generation

[0040] The data sender D1 uses a random generator to generate a random image I1. Specifically, we define the random generation space of I1 as [-2 n-1 ,2 n-1 -1], where n≥8 is a security parameter, because when n≥8, the generated image I1 is completely unrecognizable. Therefore, we take n=8 to randomly generate image I1, and then subtract I1 from the original image I to get I2.

[0041] Step 2: Key Generation

[0042] Combining Henon mapping and Sine mapping, a new efficient three-dimensional chaotic mapping is constructed as shown in formula (a):

[0043]

[0044] Assume that the image size is (ψ1, ψ2, ψ3). In order to make the constructed chaotic map in a chaotic state, we set its control parameter tuple (a, b, c) to (12.56, 0.87, 1). Use the input (X0, Y0, Z0) = (0.5, 0.5, 0.5) to iterate the chaotic map in formula (a) ψ1ψ2ψ3 times, and obtain three chaotic sequences X, Y and Z with lengths of ψ1ψ2ψ3. In order to make the chaotic sequence we obtain periodic, we use the sine function to process the above sequence to obtain a chaotic sequence with a longer period in the interval [0,1]. In order to make the random characteristics more ideal and confuse the image pixel values, the sequence values ​​are further confused to [0,255]. and Among them, Yanzhong 10 16 To extend the period, round(·) is the operation to get the nearest integer. The values ​​of the δ1 and δ2 sequences are approximately randomly distributed in the pixel interval [0, 255], so we use the δ1 and δ2 sequences as the initial encryption keys.

[0045] Step 3: Image encryption

[0046] We first stretch the images I1 and I2 sent to edge servers S1 and S2 into stream sequences of length ψ1·ψ2·ψ3, and then perform two rounds of encryption in sequence, as shown in equations (b) and (c):

[0047]

[0048] P i (i=1,2) is the encryption sequence after the first round of encryption, B i (i=1,2) is the encryption sequence after the second round of encryption.

[0049] Step 4: Image decryption

[0050] After edge servers S1 and S2 receive the encrypted images E1 and E2, they first need to perform a reverse encryption operation, where the first round of decryption is the reverse operation of the second round of encryption, and the second round of decryption is the reverse operation of the first round of encryption. The third-party server T provides decryption keys δ1 and δ2. After decryption, images I1 and I2 are obtained. The decryption operation is shown in equations (d) and (e).

[0051]

[0052] (3) Secure computing protocol design

[0053] Linear calculation protocol:

[0054] Protocol 1: Security Addition Protocol SecAdd

[0055] The data sender D1 inputs the array (u,v)∈Z n It is split into (u1, v1) and (u2, v2) and sent to edge servers S1 and S2 respectively. S1 and S2 independently calculate the sharding results f1=u1+v1 and f2=u2+v2 respectively, and then send them to the data receiver D2. D2 only needs to add the data to get f1+f2=u+v.

[0056] Protocol 2: Secure Subtraction Protocol SecSub

[0057] Similar to the secure addition protocol, you only need to change the addition into subtraction to get the calculation result.

[0058] Protocol 3: Secure Scalar Multiplication Protocol SecSMul

[0059] The data sender D1 inputs the array (p,q)∈Z n It is split into (p1, q) and (p2, q) and sent to edge servers S1 and S2 respectively. S1 and S2 independently calculate the sharding results f1=p1·q and f2=p2·q, and then send them to the data receiver D2. D2 only needs to add the data to get f1+f2=p·q.

[0060] Nonlinear Computational Protocols:

[0061] Protocol 4: Secure XOR Protocol SecXOR

[0062] Input: S i (i=1,2) has u i ∈Z n

[0063] Output: S i Return f i

[0064] 1)T randomly splits u1←ε1^ε2,u2←ο1^ο2;

[0065] 2) T will ε i and i Send to S i ;

[0066] 3) S1 calculates f1←ε1∧ο1, and S2 calculates f2←ε2∧ο2;

[0067] 4) S i Return f i .

[0068] Protocol 5: Secure Transformation From AND to XOR (STAX)

[0069] Input: Si (i=1,2) has u i ∈Z n

[0070] Output: S i Return f i

[0071] 1)T randomly generates λ i ,μ i , calculate λ←λ1∧λ2,μ←μ1∧μ2, randomly split u1←α1∧α2,u2←β1∧β2;

[0072] 2)T calculates η←λ∧μ, randomly splits η←η0∧η1, and converts λ i ,μ i ,α i ,β i ,η i Send to S i ;

[0073] 3) S i Calculate r i ←α i ∧λ i ,δ i ←β i ∧μ i ,r←r1∧r2,δ←δ1∧δ2, r i ,δ i Send to S 3-i ;

[0074] 4)S1 calculates f1←(α1&δ)∧(β1&r)∧η1

[0075] 5) S2 calculates f2←(r&δ)∧(α2&δ)∧(β2&r)∧η2;

[0076] 6) S i Return f i .

[0077] Protocol 6: Secure Transformation From XOR to Addition (STXA)

[0078] Input: S i (i=1,2) has u i ∈Z n

[0079] Output: S i Return f i

[0080] 1) S1 and S2 computing

[0081] 2) S i Compute locally

[0082] 3) S i Return f i .

[0083] Protocol 7 Security Comparison Protocol SecComp

[0084] Input: S i (i=1,2) has u i ∈Z n

[0085] Output: S i Return f i (l-1)

[0086] 1) for j = 0, · · · , l - 1 do

[0087] 2) S i Compute locally

[0088] 3) S i Compute locally and

[0089] 4) for j = 1, · · · , l - 2 do

[0090] 5) S1 and S2 computing and

[0091] 6) S i Compute locally

[0092] 7) for j = 1, · · ·, l - 1 do

[0093] 8) S1 and S2 computing

[0094] 9) S i Return f i (l-1) .

[0095] Protocol 8: Security Activation Protocol SecReLU

[0096] Input: S i (i=1,2) has input feature x i ∈Z n

[0097] Output: S iReturns the activated feature y i

[0098] 1) S1 collaborates with S2 to calculate (ω1,ω2)←SecComp(x1,x2);

[0099] 2) S1 collaborates with S2 to calculate (z1,z2)←STXA(ω1,ω2);

[0100] 3) S1 cooperates with S2 to calculate (y1,y2)←SecMul(x1,1-z1,x2,-z2);

[0101] 4) S i Return y i .

Claims

1. A lightweight privacy-preserving target classification method based on additive secret sharing, characterized by: An encryption scheme based on Henon and Sine chaotic mapping is used to encrypt the data and upload it to the edge server. After the edge server decrypts the data, the neural network is trained through an efficient and secure computing protocol based on additive secret sharing to implement a classification framework that protects data privacy.

2. According to claim 1, the lightweight privacy protection target classification method based on additive secret sharing is characterized in that The method comprises the following steps: Step 1: generating an initial image, the data sender D1 generates a random image I1 using a random generator, and obtains I2 by subtracting I1 from the original image I; Step 2: Key generation. Combining Henon mapping and Sine mapping, a new efficient three-dimensional chaotic mapping is constructed as shown in formula (a): Assume that the image size is (ψ1, ψ2, ψ3). In order to make the constructed chaotic map in a chaotic state, set the value of the control parameter tuple (a, b, c), input (X0, Y0, Z0) to iterate the chaotic map in formula (a) ψ1ψ2ψ3 times, and obtain three chaotic sequences X, Y and Z with lengths of ψ1ψ2ψ3. In order to make the chaotic sequence we obtain periodic, use the sine function to process the above sequence to obtain a chaotic sequence with a longer period in the interval [0,1]. and In order to make the random characteristics more ideal and confuse the image pixel values, the sequence values ​​are further confused and the transformation is performed. and Among them, Yanzhong 10 16 To extend the period, round(·) is the operation to the nearest integer; Step 3: Image encryption We first stretch the images I1 and I2 sent to edge servers S1 and S2 into stream sequences of length ψ1·ψ2·ψ3, and then perform two rounds of encryption in sequence, as shown in equations (b) and (c): P i (i=1,2) is the encryption sequence after the first round of encryption, B i (i=1,2) is the encryption sequence after the second round of encryption; Step 4: Image Decryption After edge servers S1 and S2 receive the encrypted images E1 and E2, they first need to perform a reverse encryption operation, where the first round of decryption is the reverse operation of the second round of encryption, and the second round of decryption is the reverse operation of the first round of encryption. The third-party server T provides decryption keys δ1 and δ2. After decryption, images I1 and I2 are obtained. The decryption operation is shown in equations (d) and (e).

3. According to claim 1, the lightweight privacy protection target classification method based on additive secret sharing is characterized in that The linear computing protocols of the secure computing protocol include: secure addition protocol, secure subtraction protocol, and secure scalar multiplication protocol.

4. According to claim 1, the lightweight privacy protection target classification method based on additive secret sharing is characterized in that The nonlinear computing protocols of the secure computing protocol include: secure XOR protocol, secure conversion protocol STAX, secure conversion protocol STXA, secure comparison protocol, and secure activation protocol.

Images