基于个性化差分隐私的联邦学习方法

Federated Learning Approach Based on Personalized Differential Privacy

联邦学习的模型参数可能导致用户隐私信息泄露,为解决此问题,将联邦学习和差分隐私进行结合的方法被广泛使用,但目前大多数方法只关注统一的隐私预算分配,忽略了由于用户数据分布不平衡带来的不同隐私预算需求。针对上述问题,提出一种基于数据分布的个性化差分隐私的联邦学习方法,根据用户间数据分布的差异,提出一种基于信息熵的隐私预算分配方案,依据信息熵为用户分配不同的隐私预算,信息熵越大的用户分配的隐私预算越高,从而量化了用户的隐私需求,实现对用户隐私的个性化保护。实验结果表明,在数据分布不平衡的场景下,相比基于统一的隐私预算分配方法,基于信息熵的隐私预算分配方法的模型准确率在不同的隐私预算下均有提高。

Federated learning model parameters may lead to the leakage of user privacy information.To address this issue,methods that integrate federated learning with differential privacy were widely used.However,most current methods focus solely on a uniform privacy budget allocation,overlooking the diverse privacy budget needs arising from imbalanced user data distributions.In response to these challenges,A federated learning approach that utilizes personalized differential privacy based on data distribution was proposed.Given the imbalanced nature of data distribution and the variance across different users,A privacy budget allocation method grounded on information entropy was introduced.Based on the differences in information entropy among users,different privacy budgets were allocated.Users with higher information entropy receive a higher privacy budget,thereby quantifying individual user's privacy needs and achieving personalized privacy protection.Experimental results show that,in scenarios with imbalanced data distributions,the model accuracy of the entropy-based privacy budget allocation method improves across various privacy budgets compared to methods using a uniform privacy budget.