Comment on “Federated Learning With Differential Privacy: Algorithms and Performance Analysis”
Krishnan Rajkumar, Antriksh Goswami, Kiruthika Sri Lakshmanan, Ruchir Gupta · IEEE Transactions on Information Forensics and Security · 2022
The research paper [1] proposes a differential privacy algorithm in the context of Federated Learning and provides its performance analysis, mainly focusing on proving a convergence bound for the loss function. In this paper, we show that some of the mathematical derivations given in [1] are not valid. Thus the bounds they prove in the paper do not hold for all loss functions. In this work, we give the correct derivation of the best possible local sensitivity bound, which is valid for all loss functions. We also state the modifications in the bounds for global sensitivity and the standard deviation of the Gaussian noise added, both before and after aggregation.