基于布隆过滤器的联邦遗忘学习

Federated Unlearning Based on Bloom Filter

联邦学习体现了集中数据收集和最小化的原则,可以有效解决传统集中式机器学习存在的数据孤岛以及系统性的隐私风险问题,但数据使用中仍存在隐私泄露风险。为此,提出一种基于布隆过滤器的联邦遗忘学习方案,通过布隆过滤器实现联邦遗忘学习的成员资格验证、用户隐私保护和数据传输完整性的验证,通过组合哈希函数将数据映射到布隆过滤器中,即使布隆过滤器被泄露,攻击者也只能获取到位数组,隐私信息仍然得到有效保护;同时,使用布隆过滤器能够快速验证数据完整性,确保传输和存储的数据未被篡改,从而提高系统的安全性。安全性分析结果表明,所提方案能够抵抗内部敌手、外部敌手和内外合谋对布隆过滤器存储数据的攻击。性能评估结果表明,在成员资格验证方面,相比于使用哈希表,布隆过滤器在插入速度上提升了约14.6%,在查找速度上提升了约5.3%。

Federated learning embodies the principle of centralized data collection and minimization,which can effectively solve the data silos and systemic privacy risk problems that exist in conventional centralized machine learning,but there is still a risk of privacy disclosure in data use.Therefore,this paper proposes a federated unlearning scheme based on Bloom filter,which achieves membership verification,user privacy protection and data transmission integrity verification of federated unlearning through Bloom filters,and maps the data into Bloom filters by combining hash functions,so that even if the Bloom filter are compromised,the attacker can only obtain the stored data,and the privacy information can still be effectively protected.At the same time,the use of Bloom filters can quickly verify the data integrity and ensure that transmitted and stored data has not been tampered with,thus increasing the security of the system.The security analysis results show that the proposed scheme can resist the attacks of internal adversaries,external adversaries and internal and external collusion on the data stored in Bloom filter.The performance evaluation results indicate that the Bloom filter improves the insertion speed by approximately 14.6%and the lookup speed by approximately 5.3%compared to using a hash table for membership verification.