Block chain assisted matrix decomposition recommendation system

Abstract

The invention discloses a block chain assisted matrix decomposition recommendation system, and relates to the technical field of block chain big data. The blockchain-assisted matrix decomposition recommendation system provided by the invention comprises a data sharing module and a model training module which are respectively deployed in a main blockchain and a side blockchain; in the data sharing module, a project potential feature matrix is obtained through a singular value decomposition algorithm with bias terms, and homomorphic encryption is carried out to store the project potential feature matrix in an interstellar file system; through a smart contract and attribute encryption method, data transmission between a sharer and a trainer is realized, so that the trainer obtains an address of an encrypted project potential feature matrix and downloads the address; in the model training module, with the help of a sharer, a trainer completes model training, and an intelligent contract is adopted to decide awards or punishment for behaviors of participants. Therefore, by the adoption of the method, a safe and transparent data sharing mechanism and a model training environment can be provided, and safe and fair data sharing is achieved.

Description

Technical Field

[0001] The present invention relates to the field of blockchain big data technology, and in particular to a blockchain-assisted matrix decomposition recommendation system. Background Art

[0002] With the rapid development of technology and the deepening of global digital transformation, the rapid growth of data has become a social phenomenon that cannot be ignored. These data span a wide range of industries and not only contain traces of user behavior and preference information, but also contain deep insights into market trends and social dynamics. However, data itself is not directly equivalent to value. Its potential insights and application potential are often hidden under the surface data and need to be explored. The recommendation system can transform massive data into personalized recommendation information that is meaningful to both users and enterprises through highly intelligent data processing and analysis, realizing the deep mining and effective utilization of data value and maximizing data value.

[0003] However, how to solve the privacy protection problem in the recommendation system is a top priority. Traditional content-based recommendation systems need to collect as much user personal information and interaction information as possible in order to provide more reasonable recommendations, which not only brings greater privacy leakage risks, but also places higher requirements on computing resources. In contrast, collaborative filtering recommendation systems rely more on users' evaluation or behavior patterns of products, and can better adapt to the increase of new users and new projects. In particular, model-based collaborative filtering methods make recommendations by learning the potential features of users and projects, and have better generalization capabilities. The collaborative filtering algorithm based on matrix decomposition obtains the preference information of potential users and the feature information of products through dimensionality reduction, and then restores the original user-product matrix through the dot product method, supplements the missing values ​​in the original sparse rating matrix, and realizes the prediction of recommendation results. In practical applications, the user rating matrix often exhibits extremely high sparsity, which seriously affects the recommendation performance of matrix decomposition. To this end, a study proposed a matrix decomposition based on federated learning, attempting to apply federated learning to the recommendation system, alleviating the privacy issues of the recommendation system by assigning model training tasks to the user's local device, and realizing the integration of data distributed throughout the network to build a more global and rich model. However, it ignores the fact that the server can infer the user's rating information for an item from the received gradient. In addition, the latent feature matrix of the shared item is aggregating heterogeneous data from multiple parts while using homomorphic encryption and differential privacy techniques to prevent personal privacy leakage. Unfortunately, the fairness and reliability of the sharing process are not considered.

[0004] In summary, the recommendation system relies heavily on user-generated data to improve the quality of personalized recommendations, and existing technologies do not design a reasonable incentive mechanism to encourage users to contribute data. Therefore, there is an urgent need for a recommendation system that can provide a secure and transparent data sharing mechanism and model training environment to provide a reference for efficient, secure, and fair data sharing and mining. Summary of the invention

[0005] The purpose of the present invention is to provide a blockchain-assisted matrix decomposition recommendation system that can achieve safe and fair data sharing and model training processes.

[0006] To achieve the above objectives, the present invention provides a blockchain-assisted matrix decomposition recommendation system, comprising:

[0007] Data sharing module, used to share the potential feature matrix of the project and perform homomorphic encryption and transmission;

[0008] The model training module supports data transmission between sharers and trainers through smart contracts, and rewards or punishes the behaviors of sharers and trainers.

[0009] Preferably, the data sharing module includes:

[0010] The sharer trains locally using the singular value decomposition algorithm with a bias term to obtain the project's potential feature matrix, which is then homomorphically encrypted and uploaded to the Interstellar File System to obtain the address.

[0011] The basic information of the sharer's payment requirements and data uploaded;

[0012] The trainer uploads the task requirements and transfers the reward and deposit to the smart contract as collateral;

[0013] The sharer requests a task and submits a task deposit. When the smart contract determines that the task requirements are met, it will be set as an auxiliary trainer to participate in subsequent tasks;

[0014] The trainer applies for the attribute private key from the sharer, encrypts his own attributes with the sharer's public key, uploads it to the main blockchain, and discloses his own public key at the same time;

[0015] The sharer calculates the trainer's attribute private key, encrypts it with the trainer's public key, and uploads it to the main blockchain;

[0016] The trainer uses the attribute private key to calculate its own hidden attributes and uploads them to the main blockchain;

[0017] When the trainer meets the access requirements, the sharer uploads the address of the encrypted project potential feature matrix;

[0018] The trainer obtains the address by decryption and downloads the homomorphically encrypted project latent feature matrix.

[0019] Preferably, the model training module includes:

[0020] The trainer uploads the homomorphically encrypted error for the sharer to access, and updates the start time parameters of decryption and the end time parameters of training;

[0021] If the trainer detects that the sharer deliberately delays decryption, it will mark it as a malicious actor, record the number of malicious acts, and punish it according to the number of malicious acts at the end of training;

[0022] The sharer uploads the error plaintext and notifies the trainer to obtain it, and updates the decryption end time parameter at the same time;

[0023] After the trainer obtains the error plaintext, he continues to train to generate intermediate parameters, uploads them to the side blockchain and notifies the sharer, and updates the start time parameters for decryption and differential privacy execution at the same time;

[0024] The sharer adds noise to the decrypted intermediate parameters, uploads them to the side blockchain, and then passes them to the trainer, and updates the end time parameters for differential privacy and the start time parameters for training;

[0025] After the trainer obtains the intermediate parameters with added noise, he continues training to generate the intermediate parameters again and conducts the next round of training.

[0026] Preferably, the trainer and the sharer publish the malicious information during the training process, and reward or punish the trainer and the sharer based on the malicious information.

[0027] Preferably, it includes utilizing a time checking algorithm to determine whether there is intentional delay or payment fraud among participants, and rewarding honest participants and punishing malicious participants.

[0028] Preferably, the time checking algorithm includes checking whether the decryption of the sharer has timed out and whether the trainer has deliberately delayed the training.

[0029] Preferably, the encrypted project latent feature matrix and model training parameters are stored in the Interstellar File System.

[0030] Therefore, the present invention adopts the above-mentioned blockchain-assisted matrix decomposition recommendation system, which has the following technical effects:

[0031] (1) Using smart contracts and attribute encryption to achieve effective access control over data can make data transmission more secure and fair, and better promote model transfer learning to improve the accuracy and personalization of recommendations, and achieve fine-grained access control of users to data during the sharing process, which can automatically achieve secure and fair data sharing without the involvement of a trusted third party;

[0032] (2) Smart contracts can transmit parameters after homomorphic encryption or differential privacy on a high-performance side chain, and monitor and verify the training behavior of participants based on a time check algorithm to ensure that they comply with the protocol during model training, automatically reward participants who make honest contributions, and punish malicious actors, thereby achieving fair and efficient training.

[0033] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a framework diagram of a blockchain-assisted matrix decomposition recommendation system;

[0035] Figure 2 is a workflow diagram of a data sharing module in an embodiment of a blockchain-assisted matrix decomposition recommendation system;

[0036] Figure 3 is a workflow diagram of a model training module in an embodiment of a blockchain-assisted matrix decomposition recommendation system;

[0037] Figure 4 It is a blockchain-assisted matrix decomposition recommendation system embodiment, wherein each scheme has recommendation performance under different privacy settings on three data sets, wherein (a) is MAE under different privacy settings on ML100K data set, (b) is MAE under different privacy settings on ML1M data set, (c) is MAE under different privacy settings on ML10M data set, (d) is RMSE under different privacy settings on ML100K data set, (e) is RMSE under different privacy settings on ML1M data set, (f) is RMSE under different privacy settings on ML10M data set;

[0038] Figure 5 The present invention is a blockchain-assisted matrix decomposition recommendation system embodiment, wherein the recommendation performance of each scheme under different encryption algorithms on three data sets, wherein (a) is the MAE under different encryption algorithms on the ML100K data set, (b) is the MAE under different encryption algorithms on the ML1M data set, (c) is the MAE under different encryption algorithms on the ML10M data set, (d) is the RMSE under different encryption algorithms on the ML100K data set, (e) is the RMSE under different encryption algorithms on the ML1M data set, and (f) is the RMSE under different encryption algorithms on the ML10M data set;

[0039] Figure 6 is the gas consumption of a data sharing module in an embodiment of a blockchain-assisted matrix factorization recommendation system;

[0040] Figure 7 Gas consumption of a model training module in an embodiment of a blockchain-assisted matrix decomposition recommendation system. DETAILED DESCRIPTION

[0041] The present invention can be explained in more detail by the following examples. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following examples.

[0042] See also Figure 1 The present invention provides a blockchain-assisted matrix decomposition recommendation system, including a data sharing module and a model training module, which respectively realize fair sharing of the project potential feature matrix and secure training of the model. By deeply integrating blockchain technology and the recommendation system, a secure and transparent data sharing mechanism and model training environment can be provided, and the ownership of the data can be returned to the user, ensuring that the user knows how his or her data is used and can obtain direct or indirect rewards from it.

[0043] Among them, the description of entities in a blockchain-assisted matrix factorization recommendation system (TrustMF) is as follows:

[0044] Data Sharers (DS) have recommended datasets and hope to share their datasets safely and fairly to gain certain benefits and realize the value transformation of data. At the same time, they need to assist trainers in the model training process.

[0045] Trainer (DT), since the rating matrix of data trainers is imperfect, they hope to obtain the latent feature matrix of high-quality items to optimize their recommendation algorithm model.

[0046] Users, i.e., users of the recommendation platform, can evaluate the accuracy of the recommendation platform and also generate preference data.

[0047] The main chain can be a currently popular blockchain, such as Ethereum, Polygon and other public chains, which is responsible for the fair sharing of the project's potential feature matrix after homomorphic encryption and the incentive distribution after training.

[0048] The sidechain, or blockchain, has high throughput, enabling trainers and sharers to communicate quickly and complete training tasks efficiently.

[0049] The Interplanetary File System (IPFS) is used to store the encrypted project potential feature matrix and some model initialization parameters, reducing the storage pressure of the blockchain.

[0050] See also Figure 2 , the data sharing module can complete the preparation work before training, as follows:

[0051] The sharer (DS) is first fully trained locally using the singular value decomposition algorithm with a bias term, so that it can capture the inherent bias caused by user rating preferences and project features to improve prediction accuracy. Then, the project potential feature matrix P obtained through training is homomorphically encrypted and then uploaded to the InterPlanetary File System (IPFS) and obtains the address HD. This process is performed by Upload (P HE ,Z)→HD function implementation.

[0052] The sharer (DS) calls setHEP(P HE id,intro DS ,remun DS ,Ω) function, upload the data sharer’s requirements and the basic information of the shared data to obtain more tasks, thereby maximizing the value of the data and the remaining computing resources.

[0053] The trainer (DT) calls setTask(taskID, intro DT ,remun DT ,trainT,DeT,DePT,endT) upload the task requirements of this training, including task introduction, task reward, training time, decryption time, task deadline, etc.; at the same time, transfer the reward and deposit to the smart contract as a guarantee.

[0054] The sharer (DS) calls requestTask(trainer,taskID,P HE id,PK DS ) function, requests a task, and submits the task deposit. The smart contract will determine whether he meets the task requirements. If so, it will set him as an auxiliary trainer to participate in subsequent tasks.

[0055] The trainer (DT) calls requestSK(sharer,P HE id,S PKds ,PK DT ) function applies for the attribute private key from DS, encrypts its own attributes with the sharer's public key and uploads it to the main blockchain, and at the same time makes its own public key public for DS to use.

[0056] The sharer (DS) calls setSK(trainer,SK PKdt ) function, calculates the trainer’s (DT) attribute private key locally, encrypts it with DT’s public key and uploads it to the main blockchain.

[0057] The trainer (DT) calls the setAttr(Ω′) function, uses the attribute private key to calculate its own hidden attributes, and uploads them to the main blockchain for verification by the smart contract.

[0058] The sharer (DS) calls setCT(P HE id,ct) function to determine whether the trainer's (DT) access request meets the requirements; if the access requirements are met, DS uploads the encrypted IPFS address ct.

[0059] The trainer (DT) calls the Decrypt(PK,ct,SK)→HD function and decrypts to obtain the address HD.

[0060] The trainer (DT) calls the Download (HD) function to download the homomorphically encrypted latent feature matrix of the project and starts local model training with the assistance of the data sharer (DS).

[0061] See also Figure 3 The model training module uses the side chain to realize the supervision and traceability of the entire training process, and realizes the trusted communication of data between the two parties through smart contracts, reducing the training cost. At this point, the necessary preparation information before training has been completed on the main chain, and the next step is to transfer to the side chain. The specific process is as follows:

[0062] The trainer (DT) calls setHEe(iterationID,trainID,e HE ) function, used to upload the homomorphic encryption error e HE , that is, the error between the true value and the predicted value, and triggers the GetHEe() event, allowing DS to access e HE ; At the same time, DT updates the decryption start time parameter DeStartT and the training end time parameter tEndT.

[0063] The trainer (DT) calls the checkDeT(iterationID,trainI,D taskI,D1) function to check whether DS deliberately delays decryption. If found, it will mark it as a malicious actor and record the number of malicious acts. At the end of training, it will be punished according to the number of malicious acts.

[0064] The sharer (DS) calls the set_e(iterationID, trainID, e) function. DS uploads the plaintext error e and triggers the Get_e() event, notifying DT to obtain e and continue training the model to generate the homomorphically encrypted intermediate parameter eP. HE , and update the decryption end time parameter DeEndT.

[0065] After the trainer (DT) obtains e, it calls set_eHEP(iterationID,trainID,eP HE ) function, the intermediate parameter eP HEUpload to the blockchain, trigger the Get_eHEP() event to notify DS, and update the start time parameter DAStartT for decryption and differential privacy execution.

[0066] The trainer (DT) calls the checkDeT(iterationID,trainID,taskID,2) function to check whether DS has intentional delay behavior.

[0067] Sharer (DS) calls addNoise(eP) → eP L Function, add appropriate noise to the decrypted intermediate parameter eP, and then pass it to DT to get eP L Among them, eP cannot be directly passed to DT because the plaintext e has been obtained in the previous step, preventing DT from inferring the potential feature P, which protects sensitive data and ensures data availability.

[0068] The sharer (DS) calls set_LeP(iterationID,trainID,eP L ) function, eP L Upload to the side blockchain, thereby triggering the Get_LeP() event, allowing DT to obtain eP L ; At the same time, update the end time parameter DAEndT for differential privacy execution and the training start time parameter tStartT.

[0069] The sharer (DS) calls the checkTrainT(iterationID,trainID,taskID) function to check whether DT intentionally delays the training process.

[0070] Trainer (DT) gets eP L Then continue training and generate e and eP again HE , and start a new round of training.

[0071] In addition, both DS and DT can call the checkDoEvil(trainer,sharer,taskID) function to trigger the checkDoEvil() event and publish malicious information during the training process.

[0072] The checkDoEvil function includes the BAMFtransfer(evildoer,share,r trainer,r taskID)D function and the deposit(trainer,taskID) function. Among them, the BAMFtransfer function must be called after the training time expires to prevent the trainer or sharer from transferring the deposit in advance. For example, when the audit finds that the sharer has carried out a poisoning attack during the auxiliary process, but when the punishment is to be carried out, the sharer has already transferred the deposit and reward by reporting the trainer's lazy training; the deposit function must be called by the DS. After the training is over, if no one does evil, the DS can use this function to transfer the deposit and reward to its own address.

[0073] In this example, it is assumed that there are semi-honest participants who will abide by the basic rules of the protocol, but may try to exploit loopholes in the protocol to reduce their workload or gain small benefits. Therefore, in order to solve the problems of deliberate delays and fraudulent rewards among participants, a series of algorithms are designed to motivate honest participants, punish malicious participants, and urge participants to actively participate in tasks, as follows:

[0074] The time check algorithm verifyTime(taskID,taskType), the parameter taskType has three values, corresponding to the possible malicious behavior of DS or DT. When taskType is 0, the function is used to check whether the DT single round training timeout has been exceeded; when taskType is 1 or 2, the function is used to check whether the DS decryption timeout has been exceeded. Taking taskType as an example, the function can be used to determine whether DT is deliberately delayed, as follows:

[0075] First, the smart contract verifies whether the caller is DS, then confirms whether taskType is equal to 0, and then verifies whether the training time has timed out. According to the time when DS initiates the verification, there are two special cases: one is that the current round of training has ended, and DT has updated the training end variable tEndT. If tEndT-tStratT is greater than the specified training time trainT, it proves that DT's training has timed out; the other is that the current round of training has not ended or DT is deliberately delayed, resulting in tEndT not being updated. You only need to judge whether the current time currentT minus the training start time tStartT is greater than trainT to determine whether DT is deliberately delayed. The other two situations are similar to the above and are not reflected in this embodiment.

[0076] In addition, in order to achieve a fair and efficient training process, this embodiment separates the update right at the beginning of the time from the update right at the end of the time, thereby constraining the participants to actively update. According to the result of the verifyTime() function, the two check functions checkDeT() and checkTrainT() are used to mark the behavior of the participants, for example, marking them as malicious actors and recording the number of malicious acts.

[0077] Through the above method, incentives or punishments are triggered for participants according to their behavioral performance, which can effectively encourage participants to contribute high-quality data or computing resources and promote collaboration among multiple parties.

[0078] Experimental verification

[0079] In order to evaluate the effectiveness of TrustMF, the FISCO-BCOS blockchain system was deployed on the Ubuntu 18.04LTS system to simulate the TrustMF scenario, and a large number of analyses and experiments were carried out.

[0080] In the experimental verification process, three real-world rating datasets (ML100K, ML1M and ML10M) commonly used in the empirical study of collaborative filtering algorithms were selected. The dataset files include user rating information on movies, such as user ID, movie ID, movie rating and timestamp. Each dataset is divided into three parts, namely the sharer dataset, the trainer dataset and the test dataset: First, each dataset is randomly divided into five non-overlapping equal parts, with the sharer and trainer taking up two parts each, and the remaining part being the test dataset.

[0081] In the experiment, the sharer trained alone locally, and the trainer trained with the help of the sharer, and the iterations were 50. Based on a large number of pre-experiments, the learning rate was set to 0.01 and the potential dimension was 20. The experiment was repeated 10 times and the average value was taken as the final result.

[0082] (1) Recommended system performance test

[0083] a. In this embodiment, it is necessary to perform differential privacy protection on the intermediate parameters after the decryption of the sharer. Therefore, we first investigate the impact of Laplace noise on the prediction and recommendation accuracy of the TrustMF scheme, and conduct experiments on different privacy budgets.

[0084] In the comparison scheme, the baseline scheme does not perform differential privacy, and the TrustMF scheme, FCMF scheme, and RSVD scheme use different privacy budgets to achieve differential privacy. In the TrustMF scheme of this embodiment, when training the model on the sharer and the trainer, factors such as user bias and item bias are considered, which can help the model more accurately capture the user's personal rating habits and the popularity of the item itself, thereby improving the accuracy of the prediction; in the FCMF scheme, only the bias factors on the trainer side are considered; in the RSVD scheme, only the trainer's own data is used, which may lead to poor recommendations due to the existence of the cold start problem.

[0085] See also Figure 4 ,As the privacy budget increases, the recommendations become better, but after the privacy budget is larger than 2.0, although it continues to increase, the changes in the mean absolute error (MAE) and the root mean square error (RMSE) are small.,Smaller privacy budget values ​​lead to a higher degree of privacy protection,,making it more difficult for attackers to infer personal data, but also lead to lower data availability,,and smaller privacy budget values ​​may require finer-grained noise calculations,,which increases the demand on computing resources.

[0086] By comparing the recommendation effects with and without adding noise, experiments on three datasets show that adding appropriate noise can achieve both privacy protection and data availability. It can also be seen that the effects of the TrustMF scheme and the FCMF scheme are better than those of the RSVD scheme and are close to the baseline scheme, which means that during model training, with the help of the item latent feature matrix trained by the sharer, the trainer can better discover the potential relationship between the data and improve the results of the training model.

[0087] b. Set the privacy budget to 2.0 to further improve the TrustMF scheme to ensure that data is not leaked. At the same time, verify the recommended effects of no homomorphic encryption and no differential privacy (baseline), only differential privacy (TrustMF with DP), only homomorphic encryption (TrustMF with HE), and homomorphic encryption plus differential privacy (TrustMF). Please refer to Figure 5 .

[0088] It can be seen that as the number of training iterations increases, the recommendation effect gradually improves. The effect of adding only homomorphic encryption (HE) is closest to the case without any processing, and the effect of adding only differential privacy (DP) is second. Although the experimental results show that using HE and DP at the same time does lead to worse recommendation results, the fluctuation is small and the recommendation accuracy is acceptable in practical applications. It is worth mentioning that the TrustMF scheme does achieve a more complete privacy protection effect during data sharing.

[0089] Since the above experiments show that the impact of the homomorphic encryption scheme of this embodiment on the effectiveness of recommendation is within an acceptable limit, and both encryption and decryption operations are very time-consuming, only the effects of the no differential privacy and no homomorphic encryption scheme (baseline) and the differential privacy scheme only (TrustMF with DP) are compared on the ML1M and ML10M data sets. The gap between the two schemes becomes smaller as the data set grows, which proves that the recommendation system of this embodiment can ensure the recommendation accuracy of the model while protecting experimental privacy.

[0090] (2) Smart Contracts

[0091] The TrustMF solution of this embodiment uses Solidity to develop smart contracts, where Solidity is a programming language for developing smart contracts on Ethereum. At the same time, in addition to compiling and deploying contracts using only FISCO-BCOS's pythonSDK, the functions of all smart contracts were tested using Remix to measure gas values, a unit of measurement of computing resources required to perform various operations on the blockchain, where the compiler version is 0.8.18.

[0092] See also Figure 6 In the data sharing module, deploying smart contracts consumes about 3398506 gas, but this operation only needs to be performed once. In the function call, only the data sharer DS calls the data introduction function setHEP and the model trainer DT calls the task release function setTask, which require more complex storage operations on the blockchain, resulting in higher gas costs, while other function calls consume less gas.

[0093] See also Figure 7 In the model training module, the gas cost is also mainly concentrated in the contract deployment, which is about 3747336 gas. The gas cost of set_eHEP and setLeP is relatively high because the round-robin algorithm is used in the implementation process, which requires a larger storage space, but it is still within an acceptable range. The gas measurement results of the two parts fully demonstrate that the smart contract designed in this embodiment can effectively reduce transaction costs, simplify transaction processes, and achieve fair and secure data sharing without the participation of a trusted third party.

[0094] Therefore, the present invention adopts the above-mentioned blockchain-assisted matrix decomposition recommendation system, which can ensure the security of data access and the legitimacy of user behavior during data sharing and model training.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A blockchain-assisted matrix decomposition recommendation system, characterized in that: include: The data sharing module is used to obtain the potential feature matrix of the project and perform homomorphic encryption and transmission; The model training module supports data transmission between sharers and trainers through smart contracts, and rewards or punishes the behaviors of sharers and trainers.

2. A blockchain-assisted matrix decomposition recommendation system according to claim 1, characterized in that: Data sharing module, including: The sharer trains locally using a biased singular value decomposition algorithm to obtain the project's latent feature matrix, which is then homomorphically encrypted and uploaded to the InterPlanetary File System to obtain an address. Basic information about the sharer's compensation requirements and data uploaded; Trainers upload task requirements and transfer rewards and deposits to smart contracts as collateral; The sharer requests a task and submits a task deposit. When the smart contract determines that the sharer meets the task requirements, it will be set as an auxiliary trainer to participate in subsequent tasks; The trainer applies for the attribute private key from the sharer, encrypts his own attributes with the sharer's public key, and uploads it to the main blockchain, while making his own public key public. The sharer calculates the trainer's attribute private key, encrypts it with the trainer's public key, and uploads it to the main blockchain; The trainer uses the attribute private key to calculate its own hidden attributes and uploads them to the main blockchain; When the trainer meets the access requirements, the sharer uploads the address of the encrypted project potential feature matrix; The trainer obtains the address through decryption and downloads the homomorphically encrypted project latent feature matrix.

3. A blockchain-assisted matrix decomposition recommendation system according to claim 2, characterized in that: Model training module, including: The trainer uploads the homomorphically encrypted error for the sharer to access and updates the decryption start time parameters and the training end time parameters; If the trainer detects that the sharer deliberately delays decryption, it will mark it as a malicious actor, record the number of malicious acts, and punish it according to the number of malicious acts at the end of training; The sharer uploads the error plaintext and notifies the trainer to obtain it, and updates the decryption end time parameter at the same time; After the trainer obtains the error plaintext, it continues training to generate intermediate parameters, uploads them to the side blockchain and notifies the sharer, while updating the start time parameters for decryption and differential privacy execution; The sharer adds noise to the decrypted intermediate parameters, uploads them to the side blockchain, and then passes them to the trainer, and updates the end time parameters for differential privacy and the training start time parameters; After the trainer obtains the intermediate parameters with added noise, he continues training to generate the intermediate parameters again and conducts the next round of training.

4. A blockchain-assisted matrix decomposition recommendation system according to claim 3, characterized in that: Trainers and sharers publish malicious information during the training process, and reward or punish trainers and sharers based on the malicious information.

5. A blockchain-assisted matrix decomposition recommendation system according to claim 4, characterized in that: This includes using time-checking algorithms to identify participants who are intentionally delaying or defrauding others of compensation, and rewarding honest participants and punishing malicious ones.

6. A blockchain-assisted matrix decomposition recommendation system according to claim 5, characterized in that: The time checking algorithm includes checking whether the sharer's decryption has timed out and whether the trainer has deliberately delayed training.

7. A blockchain-assisted matrix decomposition recommendation system according to claim 6, characterized in that: The encrypted project latent feature matrix and model training intermediate parameters are stored in the Interstellar File System.

Images