Research On Secure And Reliable Data Collaborative Computing Method In Cross-domain Recommendation

As an effective information screening tool,recommendation systems have been widely used in many industries such as smart business,decision aid,pre-diagnosis,and smart cities.The essence is to extract user preferences based on the collected interaction data to enable accurate information filtering.However,as the number of participating users and items grows rapidly,the single-domain recommendation performance is clearly affected by data sparsity and cold-start issues.Therefore,cross-domain recommendation techniques that can alleviate the aforementioned problems have become a research hot spot.The basic idea is to fill in the missing information in the target domain and improve the recommendation performance by making correlations through the similarity of the inter-domain data.However,due to the diverse data sources and types involved in cross-domain techniques,cross-domain recommendation faces many challenges such as heterogeneous collaboration,privacy disclosure,and poisoning attacks,which mainly include:(1)Multi-correlation of data in multi-source heterogeneous data fusion leads to complex privacy disclosure mechanism.Data association diversity not only enhances the background knowledge of the attacker,but also complicates the mechanism of data privacy disclosure and makes crossdomain privacy issues more challenging.(2)The coupling between privacy protection and recommendation algorithms in a cross-domain setting is complex.In cross-domain environments,domain-managed data is usually processed anonymously before performing recommendation analysis.Therefore,the methods of privacy processing directly degrade the quality of the input data and thus affect the accuracy of the recommendation results.There is a complicated coupling between the two.(3)Security threats such as data integrity and credibility are more complex and changeable in open and dynamic application environment.Cross-domain environments are characterized by dynamic openness and decentralization,and suffer from trust crises.Currently,effective data control and user management mechanisms are lacking,making it difficult to defend against poisoning attacks by participants.The above challenges can be summarized as a win-win problem of security,trust,and recommendation accuracy for cross-domain data collaboration in personalized recommendation,which is the focus of this thesis.Focusing on issues such as efficient collaboration of multi-source data in cross-domain recommendation,secure control of private information,and trust management in cross-domain environments,the main work of this dissertation summarized as follows:1.A data collaborative computing approach for cross-domain recommendation with providing privacy protection and bias correction is proposed.To solve the problem that the data privacy disclosure mode becomes more complex due to the relevance of data in cross-domain data fusion,a deep learning collaborative recommendation method with the characteristics of bias correction and differential privacy protection is proposed.To be specific,firstly,the method corrects the outliers in the target domain through auxiliary domain information fusion to improve the accuracy of recommendation.Secondly,in view of the more complex privacy disclosure problem in the process of data fusion caused by the association of multiple data,a differential privacy model is adopted to disturb users’ score by adding noise during the collaborative training to ensure the security of privacy data.Then,as there is no metric to evaluate the performance of recommendation algorithms in the long tail effect,a new evaluation metric-discovery is proposed to measure the ability of recommendation algorithms to discover users’ implicit needs.Finally,through theoretical analysis and simulation experiment results,it is proved that the proposed method can improve the recommendation performance of F1 and discovery metrics by more than 4%,and the diversity metric performance is more stable,provided that privacy is guaranteed.2.A data collaborative computing method with secure storage and trusted traceability for cross-domain recommendation is proposed.In view of the complex coupling relationship between privacy processing mechanism and recommendation algorithm in the process of cross-domain data collaboration,we introduce blockchain technology to propose a secure and accurate cross-domain collaborative recommendation model.In detail,first of all,the model adopts multi-chain structure to realize the separation and storage of transaction data,user attributes and projects,and realize the security of privacy information without modifying the data.Secondly,the recommendation center in this model only collects the relationship chain without user or product information to realize the lossless and safe data fusion.Because the recommendation center cannot associate sensitive relationships with specific users without knowing user attributes and project information.Therefore,for the recommending center,the privacy of the data is safe.Then,in order to ensure the credibility of the data in the cross-domain context,a signature-based verifiable mechanism is designed to prevent transaction data from being tampered.Further,the mechanism ensures that recommendation analysis is based on the unmodified transaction data for guaranteeing the accuracy of the recommendation model.Meanwhile,an incentive mechanism based on contribution is designed to reduce the attack probability of malicious nodes.Finally,experimental validation on real-world datasets shows that the proposed model not only improves the precision and F1 metrics by more than 2 percent,but also keeps the degree of discovery and diversity metrics more stable while protecting the security credibility of the data.3.A data trust collaborative computing method based on multi-feature knowledge graph and blockchain for cross-domain recommendation is proposed.Aiming at the limitation that the above methods can only be applied to scenarios with overlapping users,this thesis proposes a trusted cross-domain recommendation model based on multi-feature collaborative knowledge graph and blockchain,which can be applied to different application scenarios.First,we analyze the differences between privacy features and recommendation features,and propose a feature knowledge graph suitable for heterogeneous data security representation,which can achieve complete representation of complex associated knowledge.Secondly,in order to realize the secure fusion of data in distributed environment,the federated learning framework is integrated to realize the distributed machine learning with privacy protection,and the distributed co-trust mechanism is established based on blockchain,which can realize the safe and lossless fusion and trusted management of multi-correlation data and ensure the accuracy of cross-domain recommendation.This model can be applied to the scenario where users/projects overlap or not,and can guarantee the privacy security and authenticity verification of data without the generalization of privacy protection technology and coordination of centralized server,realizing multiple guarantees of recommendation accuracy,data credibility and privacy.Finally,through theoretical analysis and simulation experiment results,it is proved that this proposed method can provide a secure and reliable guarantee for data fusion,with a clear advantage of increasing the F1 metric by more than 1.5% and the diversity metric by 18%compared to the baselines.4.A distributed data security control and user behavior management mechanism based on blockchain for cross-domain recommendation to achieve decentralized data trust management is proposed.The analysis shows that data fusion in the cross-domain recommendation scenario has the characteristics of dynamic openness and decentralization,which leads to more complex and changeable security threats such as data integrity and credibility.Although blockchain technology can break the limitations of existing authentication-based mechanisms for data security management,it is difficult to meet the high transaction rate application requirements of crossdomain recommendation due to consensus latency.Aiming at this problem,a blockchain-based distributed data security control and user behavior management mechanism is proposed to enable user behavior management and effective data verification in distributed environments.Firstly,the idea of blockchain fragmentation is adopted and RAFT protocol is used as a local consensus mechanism to improve the efficiency of consensus algorithm to meet the high transaction rate requirements of cross-domain recommendation applications.Secondly,a behavior-based incentive mechanism is designed to compensate for the fault tolerance of RAFT protocol,and a supervised local consensus policy and weighted rejoin rules is proposed to resist the stir attack in the blockchain,which can ensure the security of partitions and improve the fault tolerance of the system.Finally,through theoretical analysis and simulation experiments,the proposed protocol can effectively improve the efficiency of consensus algorithm,ensure the security of decentralized data fusion,achieve secure and trusted data management in cross-domain recommendation applications.5.The application case of secure and reliable cross-domain data collaborative computing and intelligent recommendation to meet the requirements of industrial applications is designed.This dissertation mainly focuses on the conflicts between the security of private data,the verifiability of anonymous data and the precision of recommendation results in cross-domain recommendation,and proposes a series of solutions for different scenarios.The key technologies of multi-source data collaborative computing are used in the design of system with the application requirements of related industries.We design the specific application in the field of agriculture concrete platform to verify the feasibility,effectiveness and practicability of our technologies.The studies of this thesis can provide a secure and trusted solution for data collaborative computing in cross-domain recommendation scenarios,which will promote the theoretical research of trusted verification technology in the security field and the application development of recommendation systems in distributed cross-domain environments.