Privacy-Preserving Federated Aggregation Technology For Data Quality

Data-driven deep learning systems have shown substantial improvement in practical applications such as autonomous driving,speech recognition and image processing with the rapid development of artificial intelligence(AI)technology.As AI technology represented by deep learning relies on massive data,the utilization of high-value data aggravates the risk of user privacy data leakage,posing new challenges to data industrialization and security.In such a context,federated learning technology emerges in response to handling dilemmas of data island and privacy protection.Currently,the federated learning technology confronts problems such as the computational efficiency to be improved in ciphertext framework,low-quality user data involved in the system(known as irregular users),and noise labels affecting the aggregation efficiency of the global model.Specifically,users involved in federated learning possess large numbers of noise data or labels that might lead to a decline in efficiency and accuracy of the global aggregation model or the loss of practical value in the aggregation model.Hence,it is of important theoretical significance and practical value to study privacy-preserving federated learning aggregation technology-oriented at data quality.The main research findings are shown below.(1)This thesis proposes a privacy-preserving federated learning aggregation framework for supporting most irregular users.To be concrete,a privacy-preserving federated learning framework(PPRFL)is constructed in accordance with the designed security division protocol for the robust proportion of irregular users against the problems of declined aggregation efficiency caused by most irregular users in the existing federated learning and privacy of leaked parameters using plaintext communication.Moreover,this thesis verifies the reliability of the dynamically updated model of the historical information of the loss value on the set and optimizes the convergence rate by reducing the model weight under the guidance of prior knowledge.On this basis,the security and feasibility of the verification scheme are verified through theoretical analysis and experimental results,respectively.(2)It proposes a privacy-preserving federated learning aggregation scheme for noise labels.This thesis designs a noise label detection federated aggregation scheme(ASNL)with the supervised comparative learning algorithm as its core to address the problems of low convergence efficiency in most noise label environments and difficulty in applying to the privacy-preserving architecture of the existing schemes.Furthermore,this thesis constructs a feature difference scoring mechanism operating on the client to filter noise label samples according to the difference in characteristics between difficult samples and other samples in the supervised comparative learning space.Also,the scoring weight of the user model with model discrimination is proposed in line with feature differences to ensure that the global model change is constructed by high-quality label user models by setting the weight reduction of the model.The author finally proposes a PP-ASNL scheme to meet the needs of privacy-preserving in conjunction with the PPRFL privacy-preserving framework against the problem of the existing schemes not considering the privacy of the user model.The experimental verification shows that the ASNL scheme is superior to the existing scheme in terms of convergence efficiency on the premise of ensuring the model accuracy.