Personalized Differential Privacy Client Selection: A Loss and Privacy Budget based Approach

Federated Learning (FL) is a distributed machine learning paradigm that enables collaborative model training while preserving data locality. To enhance privacy protection, many existing FL approaches incorporate differential privacy (DP) mechanisms. However, these methods often fail to account for the heterogeneity in clients’ privacy requirements, which can result in imbalanced noise impacts on the global model, thereby compromising training stability and convergence. In this work, the effects of differentiated privacy requirements across different training phases on model convergence are explored. Fed-LPCS, a personalized differential privacy client selection algorithm, is proposed, which dynamically balances local loss and privacy budgets to optimize client participation. By introducing adaptive weighting for these factors, our approach ensures more effective and stable training. Experimental results on multiple datasets demonstrate that in scenarios with heterogeneous privacy demands, our algorithm significantly improves model stability, reducing average performance fluctuations by 41.0%-69.2% compared to baseline methods, while achieving comparable accuracy.