Research On Hybrid Privacy Protection Technology Based On Vertical Federated Learning

In recent years,with the proliferation of AI-driven services including recommender systems and natural language processing,data privacy and security have attracted worldwide attention.Increasingly stringent data privacy and security requirements have become an emerging trend in laws and regulations around the world.As a relatively cutting-edge new technology,federated learning is a distributed training method that uses data sets scattered among various participants,integrates data information from multiple parties through privacy protection technology,and collaboratively builds a global model.Based on the existing research on the privacy issues of federated learning,this paper designs a method for the frequent interaction and direct exchange of intermediate results of model training between vertical federated learning participants,which may leak training data and generate a lot of computational overhead.More flexible and efficient hybrid security technology to achieve the purpose of optimizing the model while ensuring privacy.The main research results of this paper are as follows:(1)This work removes the role of the third-party coordinator,adopts the privacy set intersection protocol based on RSA blind signature,applies the hybrid differential privacy framework of vertical federated learning to analyze the sensitivity,and directly disturbs each training iteration.To obtain the intermediate results of the participants,the Taylor approximation of the loss function is not required.Moreover,the system implements parallel distributed logistic regression to process a large amount of training data.The two parties perform continuous local gradient updates and transmit intermediate results to each other on a unique communication channel.Subsequently theory proves both privacy assurance and utility analysis.Finally,the experiment verify that when a suitable privacy budget ε is set to 1,the algorithm Hybrid-VFL achieves an acceptable accuracy-privacy trade-off.In addition,adjusting the number of hyperparameter rounds e and weight constraints k will also have some impact on the accuracy,and the running time cost is compared to that of homomorphic encryption.The method is 2～3 times faster,and the AUC convergence effect on the data set is obvious,and the system is scalable,and its worker nodes are linearly accelerated.(2)This work conducts model training in the framework of a longitudinal federated learning system with a third-party coordinator and proposes a categorical cross-entropy loss function that deploys a gradient-based optimizer on the client rather than a centralized server.Then,in the training process of the federated model,the privacy protection entity alignment technology is used to obtain the common ID of both parties,and the Paillier addition homomorphic encryption scheme is introduced to encrypt and decrypt the partial gradient sent by the passive party to the active party.The security of the algorithm is based on the large integer factorization problem,and compute the exact gradient without approximation.Subsequently theory proves both performance and privacy.Finally,the experiment verify that when the number of iterations is 10～50 rounds,this the algorithm CCE-VFL can minimize the training loss to about 0.45 at a similar speed,in addition,the accuracy and precision increase to 74% and 77% at a certain speed,and the AUC area reaches about 0.75,which all show acceptable training results and also eliminate concerns about floating-point precision,it shows that the system communication cost is also reduced when the data privacy of the participants is protected.