Analysis method based on AIGC privacy protection data

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an analysis method based on AIGC privacy protection data. The method comprises the following steps: firstly, calling a weight hidden multi-user function encryption technology to encrypt AIGC privacy protection data, and sending the data to an AICG server; secondly, decrypting the encrypted AIGC privacy protection data to obtain an AIGC privacy protection aggregation model; then calling a Bayesian difference privacy technology, and according to parameter distribution characteristics of the AIGC privacy protection aggregation model, adding correspondingly distributed noise to weight parameters of the AIGC privacy protection aggregation model and correcting the noise; and finally, calculating privacy loss according to the corrected AIGC privacy protection aggregation model, and analyzing the leakage probability of the current AIGC privacy protection data. The privacy and the security of the data are ensured; noise is added and corrected according to the distribution of the model, an attacker is prevented from reasoning the model, and the model availability is ensured while privacy is effectively protected.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an analysis method based on AIGC privacy protection data. Background Art

[0002] With the continuous advancement of artificial intelligence technology, AI-generated content (AIGC) has become a hot field, involving content creation, data analysis, pattern recognition and other aspects. The development of AIGC technology has brought new productivity tools, but also triggered discussions about privacy and security. AIGC systems may encounter the risk of privacy leakage when processing large amounts of data. How to ensure the privacy and security of data has become an urgent problem to be solved. In addition, the illegal application of AIGC may cause potential harm to social, political, financial and other fields, which is also a focus that needs attention.

[0003] In order to meet these challenges, in addition to formulating corresponding laws, regulations and policies, it is also necessary to rely on technical means to protect privacy and security. Privacy-preserving computing technologies, including federated learning, secure multi-party computing, differential privacy, homomorphic encryption, trusted execution environment, etc., are key technologies to solve this problem. Globally, the development and application of AIGC technology require international cooperation and communication. By sharing research results and experience, we can jointly promote the development of privacy, security and protection technologies of AIGC systems. With the continuous advancement of technology and the expansion of application scenarios, AIGC's privacy, security and protection technologies will face more challenges and opportunities. Summary of the invention

[0004] Aiming at the problem that the AIGC system has the risk of privacy leakage when processing a large amount of data and cannot ensure the privacy and security of the data, the present invention proposes an analysis method based on AIGC privacy protection data; firstly, the weight hidden multi-user function encryption technology is called to encrypt the AIGC privacy protection data and send it to the AICG server; secondly, the encrypted AIGC privacy protection data is decrypted to obtain the AIGC privacy protection aggregation model; then, the Bayesian differential privacy technology is called, and according to the parameter distribution characteristics of the AIGC privacy protection aggregation model, the corresponding distribution of noise is added to the weight parameters of the AIGC privacy protection aggregation model and corrected; finally, the privacy loss is calculated according to the corrected AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed; thus, the privacy and security of the data are ensured.

[0005] The specific implementation contents of the present invention are as follows:

[0006] An analysis method based on AIGC privacy protection data, specifically including:

[0007] Call the weight-hidden multi-user function encryption technology to encrypt the AIGC privacy protection data obtained from the AIGC application, and send the encrypted AIGC privacy protection data to the AICG server;

[0008] Decrypting the encrypted AIGC privacy protection data according to the distribution function obtained from the key management center to obtain the AIGC privacy protection aggregation model;

[0009] Calling Bayesian differential privacy technology, according to the distribution characteristics of AIGC privacy protection aggregation model parameters, adding corresponding distribution noise to the weight parameters of AIGC privacy protection aggregation model and correcting it;

[0010] The privacy loss is calculated based on the revised AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed.

[0011] In order to better implement the present invention, further, the weight-hidden multi-user function encryption technology is called to encrypt the AIGC model parameters, and the encrypted AIGC privacy protection data is sent to the AICG server, specifically including:

[0012] Obtain AIGC privacy-protected data from the AIGC application;

[0013] Generate the master public key and master private key according to the stepup algorithm obtained from the key management center, and send the corresponding encryption key and function decryption key to the AICG server;

[0014] The AIGC privacy protection data is encrypted according to the encryption key, and the encrypted AIGC privacy protection data is sent to the AICG server.

[0015] In order to better implement the present invention, further, the encrypted AIGC privacy protection data is decrypted according to the allocation function obtained from the key management center to obtain the AIGC privacy protection aggregation model, which specifically includes:

[0016] Send the set weights to the key management center and combine them to obtain a weight vector;

[0017] Split the weight vector according to the keygen algorithm obtained from the key management center, and call the Enc algorithm to generate the function decryption key;

[0018] Call the KeyDistribute algorithm and obtain the label of AIGC privacy protection data based on the master public key and master private key;

[0019] Generate AIGC privacy-protected data public key and AIGC privacy-protected data private key based on the label;

[0020] Encrypt AIGC privacy protection data according to AIGC privacy protection data public key and generate AIGC privacy protection data ciphertext;

[0021] The decryption function Dec is called to calculate the discrete logarithm of the AIGC privacy-protected data ciphertext to obtain the AIGC privacy-protected aggregation model.

[0022] In order to better implement the present invention, further, the Bayesian differential privacy technology is called, and according to the distribution characteristics of the AIGC privacy protection aggregation model parameters, the corresponding distribution noise is added to the weight parameters of the AIGC privacy protection aggregation model and corrected, specifically including:

[0023] Calculate the expected distribution value of the parameters of the AIGC privacy-preserving aggregation model;

[0024] According to the set Gaussian mechanism noise parameters and AIGC privacy protection data sampling probability, the output distribution of the AIGC privacy protection aggregation model is calculated;

[0025] According to the output distribution, the noise of the corresponding distribution is added to the weight parameters of the AIGC privacy-preserving aggregation model;

[0026] Call Bayesian differential privacy technology to correct the noise of the corresponding distribution.

[0027] In order to better implement the present invention, further, the privacy loss is calculated according to the modified AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed, which specifically includes:

[0028] According to the output distribution and binomial distribution function, the expected value of AIGC privacy-preserving data distribution is calculated;

[0029] Calculate the privacy loss function based on the expected value of the AIGC privacy-preserving data distribution;

[0030] The total privacy loss is calculated according to the privacy loss function, and the leakage probability of the current AIGC privacy-protected data is analyzed according to the set privacy budget.

[0031] In order to better implement the present invention, further, analyzing the leakage probability of the current AIGC privacy protection data according to the set privacy budget specifically includes:

[0032] If the total privacy loss is greater than the privacy budget, the probability of leakage of the current AIGC privacy protection data is analyzed to be high, and the federated learning AIGC model training is stopped;

[0033] If the total privacy loss is greater than the privacy budget, the leakage probability of the current AIGC privacy protection data is analyzed to see if it meets the requirements, and the iterative training of the federated learning AIGC model continues.

[0034] The present invention has the following beneficial effects:

[0035] (1) The present invention adopts Bayesian differential privacy to analyze AIGC privacy protection data in federated learning, analyzes the leakage probability of current AIGC privacy protection data, and ensures the security and privacy of AIGC privacy protection data.

[0036] (2) The present invention adds noise and makes corrections based on the distribution of the model to prevent attackers from reasoning about the model, effectively protecting privacy while ensuring model availability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flowchart of the analysis method based on AIGC privacy protection data provided by the present invention. DETAILED DESCRIPTION

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be understood that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, and therefore should not be regarded as limiting the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical personnel in this field without making creative work are within the scope of protection of the present invention.

[0039] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "disposed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0040] Embodiment 1:

[0041] This embodiment proposes an analysis method based on AIGC privacy protection data, which specifically includes:

[0042] Call the weight-hidden multi-user function encryption technology to encrypt the AIGC privacy protection data obtained from the AIGC application, and send the encrypted AIGC privacy protection data to the AICG server;

[0043] Decrypting the encrypted AIGC privacy protection data according to the distribution function obtained from the key management center to obtain the AIGC privacy protection aggregation model;

[0044] Calling Bayesian differential privacy technology, according to the distribution characteristics of AIGC privacy protection aggregation model parameters, adding corresponding distribution noise to the weight parameters of AIGC privacy protection aggregation model and correcting it;

[0045] The privacy loss is calculated based on the revised AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed.

[0046] Furthermore, the weight-hidden multi-user function encryption technology is called to encrypt the AIGC model parameters, and the encrypted AIGC privacy protection data is sent to the AICG server, specifically including:

[0047] Obtain AIGC privacy-protected data from the AIGC application;

[0048] Generate the master public key and master private key according to the stepup algorithm obtained from the key management center, and send the corresponding encryption key and function decryption key to the AICG server;

[0049] The AIGC privacy protection data is encrypted according to the encryption key, and the encrypted AIGC privacy protection data is sent to the AICG server.

[0050] Furthermore, decrypting the encrypted AIGC privacy protection data according to the allocation function obtained from the key management center to obtain the AIGC privacy protection aggregation model specifically includes:

[0051] Send the set weights to the key management center and combine them to obtain a weight vector;

[0052] Split the weight vector according to the keygen algorithm obtained from the key management center, and call the Enc algorithm to generate the function decryption key;

[0053] Call the KeyDistribute algorithm and obtain the label of AIGC privacy protection data based on the master public key and master private key;

[0054] Generate AIGC privacy-protected data public key and AIGC privacy-protected data private key based on the label;

[0055] Encrypt AIGC privacy protection data according to AIGC privacy protection data public key and generate AIGC privacy protection data ciphertext;

[0056] The decryption function Dec is called to calculate the discrete logarithm of the AIGC privacy-protected data ciphertext to obtain the AIGC privacy-protected aggregation model.

[0057] Furthermore, the Bayesian differential privacy technology is called to add and correct the corresponding distributed noise to the weight parameters of the AIGC privacy protection aggregation model according to the distribution characteristics of the AIGC privacy protection aggregation model parameters, specifically including:

[0058] Calculate the expected distribution value of the parameters of the AIGC privacy-preserving aggregation model;

[0059] According to the set Gaussian mechanism noise parameters and AIGC privacy protection data sampling probability, the output distribution of the AIGC privacy protection aggregation model is calculated;

[0060] According to the output distribution, the noise of the corresponding distribution is added to the weight parameters of the AIGC privacy-preserving aggregation model;

[0061] Call Bayesian differential privacy technology to correct the noise of the corresponding distribution.

[0062] Furthermore, the privacy loss is calculated according to the modified AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed, which specifically includes:

[0063] According to the output distribution and binomial distribution function, the expected value of AIGC privacy-preserving data distribution is calculated;

[0064] Calculate the privacy loss function based on the expected value of the AIGC privacy-preserving data distribution;

[0065] The total privacy loss is calculated according to the privacy loss function, and the leakage probability of the current AIGC privacy-protected data is analyzed according to the set privacy budget.

[0066] Furthermore, the analysis of the leakage probability of the current AIGC privacy protection data according to the set privacy budget specifically includes:

[0067] If the total privacy loss is greater than the privacy budget, the probability of leakage of the current AIGC privacy protection data is analyzed to be high, and the federated learning AIGC model training is stopped;

[0068] If the total privacy loss is greater than the privacy budget, the leakage probability of the current AIGC privacy protection data is analyzed to see if it meets the requirements, and the iterative training of the federated learning AIGC model continues.

[0069] Working principle: This embodiment first calls the weight-hidden multi-user function encryption technology to encrypt the AIGC privacy protection data and send it to the AICG server; secondly, it decrypts the encrypted AIGC privacy protection data to obtain the AIGC privacy protection aggregation model; then it calls the Bayesian differential privacy technology, and according to the parameter distribution characteristics of the AIGC privacy protection aggregation model, adds the corresponding distribution noise to the weight parameters of the AIGC privacy protection aggregation model and corrects it; finally, it calculates the privacy loss based on the corrected AIGC privacy protection aggregation model, and analyzes the current probability of leakage of AIGC privacy protection data; thus, the privacy and security of the data are ensured.

[0070] Embodiment 2:

[0071] This embodiment is based on the above embodiment 1. Figure 1 As shown, a specific embodiment is used for description.

[0072] Step S1: Call the weight-hidden multi-user function encryption technology to encrypt the AIGC privacy protection data obtained from the AIGC application, and send the encrypted AIGC privacy protection data to the AICG server;

[0073] Step S11: Obtain AIGC privacy protection data from the AIGC application;

[0074] Step S12: Generate the master public key zg and the master private key zs according to the stepup algorithm obtained from the key management center, and send the corresponding encryption key and function decryption key to the AICG server;

[0075] Step S13: Encrypt the AIGC privacy protection data according to the encryption key, and send the encrypted AIGC privacy protection data to the AICG server.

[0076] Step S2: decrypting the encrypted AIGC privacy protection data according to the distribution function obtained from the key management center to obtain the AIGC privacy protection aggregation model;

[0077] Step S21: Send the set weight w to the key management center and combine to obtain the weight vector Q;

[0078] Step S22: split the weight vector Q according to the keygen algorithm obtained from the key management center, and call the Enc algorithm to generate a function decryption key;

[0079] Step S23: Call the KeyDistribute algorithm and obtain the label i of the AIGC privacy protection data according to the master public key zg and the master private key zs;

[0080] Step S24: Generate AIGC privacy protection data public key zg according to label i i、AIGC privacy protection data private key zs i ;

[0081] Step S25: Encrypt the AIGC privacy protection data according to the AIGC privacy protection data public key and generate the AIGC privacy protection data ciphertext C i m ;

[0082] Step S26: Call the decryption function Dec to calculate the discrete logarithm of the AIGC privacy protection data ciphertext, and obtain the AIGC privacy protection aggregation model based on the weighted average of the discrete logarithm.

[0083] Step S3: Call the Bayesian differential privacy technology, and add the corresponding distributed noise to the weight parameters of the AIGC privacy protection aggregation model and correct it according to the parameter distribution characteristics of the AIGC privacy protection aggregation model;

[0084] Step S31: Calculate the distribution expectation value E of the AIGC privacy protection aggregation model parameters;

[0085] Step S32: Calculate the output distribution of the AIGC privacy protection aggregation model according to the set Gaussian mechanism noise parameters and AIGC privacy protection data sampling probability. The specific process is as follows:

[0086] y(w (t) |w (t-1) ,D')=(1-n)N(g t ,σ 2 )+nN(g t ',σ 2 )

[0087] Among them, σ is the noise parameter of the Gaussian mechanism, n is the sampling probability of AIGC privacy protection data, and g t is the non-private output of the t-th iteration that does not contain the AIGC privacy-preserving data sample, g t ' is the non-privacy output of the tth iteration containing AIGC privacy-preserving data samples, D' is the AIGC privacy-preserving dataset, w (t) is the random output of the tth iteration.

[0088] Step S33: according to the output distribution, adding noise of the corresponding distribution to the weight parameters of the AIGC privacy protection aggregation model;

[0089] This embodiment adds noise of corresponding distribution to the weight parameters of the AIGC privacy protection aggregation model to prevent attackers from obtaining accurate output results by querying the model, and also prevents attackers from inferring whether the current data belongs to AIGC privacy protection data from the output results.

[0090] Step S34: Call the Bayesian differential privacy technology to correct the noise of the corresponding distribution.

[0091] Step S4: Calculate the privacy loss based on the modified AIGC privacy protection aggregation model and analyze the leakage probability of the current AIGC privacy protection data.

[0092] Step S41: Calculate the expected value of the AIGC privacy protection data distribution according to the output distribution and the binomial distribution function B(λ+1,q);

[0093] Among them, λ+1 is the number of experiments of binomial distribution, and q is the probability of success of binomial distribution experiment.

[0094] Step S42: Calculate the privacy loss function according to the expected value of the AIGC privacy protection data distribution;

[0095]

[0096] Step S43: Calculate the total privacy loss according to the privacy loss function, and analyze the leakage probability of the current AIGC privacy protection data according to the set privacy budget.

[0097] Step S431: If the total privacy loss is greater than the privacy budget, the probability of leakage of the current AIGC privacy protection data is analyzed to be high, and the federated learning AIGC model training is stopped;

[0098] Step S432: If the total privacy loss is greater than the privacy budget, the leakage probability of the current AIGC privacy protection data is analyzed to see if it meets the requirements, and the iterative training of the federated learning AIGC model is continued.

[0099] The other parts of this embodiment are the same as those of the above-mentioned embodiment 1, and thus will not be described in detail.

[0100] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An analysis method based on AIGC privacy protection data, characterized in that: include: Call the multi-user function encryption method to encrypt the AIGC privacy protection data obtained from the AIGC application, and send the encrypted AIGC privacy protection data to the AICG server; According to the distribution function obtained from the key management center, the encrypted AIGC privacy protection data is decrypted through the distribution function to obtain the AIGC privacy protection aggregation model; According to the parameter distribution characteristics of the AIGC privacy protection aggregation model, the Bayesian differential privacy method is called to add the corresponding distribution noise to the weight parameters of the AIGC privacy protection aggregation model and correct it; The privacy loss is calculated based on the revised AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed.

2. According to claim 1, the analysis method based on AIGC privacy protection data is characterized in that: The encryption of AIGC privacy protection data by the multi-user function encryption method specifically includes: Obtain AIGC privacy-protected data from the AIGC application; Generate the master public key and master private key according to the stepup algorithm obtained from the key management center, and send the corresponding encryption key to the AICG server; Encrypt AIGC privacy protection data according to the encryption key.

3. According to claim 2, the analysis method based on AIGC privacy protection data is characterized in that: The AIGC privacy protection aggregation model is obtained according to the distribution function obtained from the key management center and the encrypted AIGC privacy protection data is decrypted by the distribution function, specifically including: Send the set weights to the key management center and combine them to obtain a weight vector; Split the weight vector according to the keygen algorithm obtained from the key management center, and call the Enc algorithm to generate the function decryption key; Call the KeyDistribute algorithm and obtain the label of AIGC privacy protection data based on the master public key and master private key; Generate AIGC privacy-protected data public key and AIGC privacy-protected data private key based on the label; Encrypt AIGC privacy protection data according to AIGC privacy protection data public key and generate AIGC privacy protection data ciphertext; The decryption function Dec is called to calculate the discrete logarithm of the AIGC privacy-protected data ciphertext to obtain the AIGC privacy-protected aggregation model.

4. According to claim 3, the analysis method based on AIGC privacy protection data is characterized in that: According to the parameter distribution characteristics of the AIGC privacy protection aggregation model, the Bayesian differential privacy method is called to add corresponding distribution noise to the weight parameters of the AIGC privacy protection aggregation model and correct it, specifically including: Calculate the expected distribution value of the parameters of the AIGC privacy-preserving aggregation model; According to the set Gaussian mechanism noise parameters and AIGC privacy protection data sampling probability, the output distribution of the AIGC privacy protection aggregation model is calculated; According to the output distribution, the noise of the corresponding distribution is added to the weight parameters of the AIGC privacy-preserving aggregation model; Call Bayesian differential privacy technology to correct the noise of the corresponding distribution.

5. According to claim 4, the analysis method based on AIGC privacy protection data is characterized in that: The privacy loss is calculated according to the modified AIGC privacy protection aggregation model, and the leakage probability of the current AIGC privacy protection data is analyzed, specifically including: According to the output distribution and binomial distribution function, the expected value of AIGC privacy-preserving data distribution is calculated; Calculate the privacy loss function based on the expected value of the AIGC privacy-preserving data distribution; The total privacy loss is calculated according to the privacy loss function, and the leakage probability of the current AIGC privacy-protected data is analyzed according to the set privacy budget.

6. The analysis method based on AIGC privacy protection data according to claim 5 is characterized in that: The analysis of the leakage probability of the current AIGC privacy protection data according to the set privacy budget specifically includes: If the total privacy loss is greater than the privacy budget, the probability of leakage of the current AIGC privacy protection data is analyzed to be high, and the federated learning AIGC model training is stopped; If the total privacy loss is greater than the privacy budget, the leakage probability of the current AIGC privacy protection data is analyzed to see if it meets the requirements, and the iterative training of the federated learning AIGC model continues.

7. The analysis method based on AIGC privacy protection data according to claim 1 is characterized in that: The multi-user function encryption method is a weight-hidden multi-user function encryption method.

Images