联邦遗忘学习隐私安全与算法效率研究综述

With the rapid advancement of data-driven artificial intelligence applications, concerns regarding personal data security and privacy protection have intensified among users. Federated Learning (FL) effectively mitigates issues such as privacy breaches and insufficient computational resources in traditional centralized model training modes. While legislations such as the EU General Data Protection Regulation and the California Consumer Privacy Act require ensuring both data security and participants "right to be forgotten", FL participants may request the forgetting of their own data contributions to protect their privacy, and FL systems may intend to delete contributions from malicious users to maintain robustness. Federated Unlearning (FU) technologies address these forgetting requests, ensuring utility restoration of the original model while performing forgetting operations on participants or data samples, which has become a hot topic in current research. However, existing methods still have shortcomings in ensuring privacy security and improving algorithm efficiency. This paper first systematically introduces the basic concepts of federated unlearning and highlights the four core challenges it faces: high risk of privacy leakage, difficulty in restoring model performance, high computational overhead, and high storage costs. Subsequently, it provides a comprehensive review of research progress in federated unlearning from four aspects: privacy protection, model recovery, computational efficiency, and storage efficiency. Furthermore, the paper categorizes and compares related approaches clearly to provide a solid theoretical and practical foundation for future research. Finally, the paper summarizes the practical applications of federated unlearning and discusses potential future research directions to promote its safe application in the field of artificial intelligence. Compared with the existing Federated Unlearning survey, this paper differs in five aspects: (1) This paper identifies the core threats in federated unlearning through a thorough analysis of existing literature and approaches. These threats include privacy leakage, difficulty in model recovery, high computational costs, and increased storage costs. Unlike other reviews that mainly provide an overview, this paper discusses the limitations of various approaches in tackling these challenges and examines their specific impacts on model utility, system security, and algorithm efficiency. (2) This paper systematically summarizes research on privacy protection in federated unlearning, analyzing the effectiveness, applicability, and limitations of various strategies from multiple perspectives. It delves into differential privacy, homomorphic encryption, and other protocols, highlighting their effectiveness in safeguarding user data privacy and addressing privacy leakage. (3) This paper explores how to restore model performance through measures such as clearing residual data, verifying forgetting results, and enhancing model robustness. By comparing various data removal and verification strategies in terms of their effectiveness in thoroughly eliminating residual forgotten data, preventing catastrophic federated unlearning, and ensuring model stability, it reveals their strengths and weaknesses. (4) This paper reviews and evaluates methods for improving the computational and storage efficiency of federated unlearning. It discusses specific techniques such as compression strategies, sharding mechanisms, and knowledge distillation, comparing their advantages and disadvantages regarding speed, computational complexity, and space complexity. (5) This paper summarizes practical application scenarios of federated unlearning found in the literature, including personalized recommendations, healthcare, digital twins, and ranking learning. By organizing and analyzing these scenarios, it demonstrates the applicability, effectiveness, and limitations of federated unlearning in real-world contexts, providing a foundation for scholars to further explore solutions in specific scenarios.