Differential privacy mechanisms in federated learning: privacy-utility trade-offs and future directions

Federated learning (FL) has become another approach to protect privacy by training models without transmitting the data on which the models work. Yet, it is established that FL remains susceptible to certain privacy concerns like membership inference and model inversion. This is due to the reason that FL has incorporated with Differential Privacy (DP), which is used to help reduce these risks through injecting carefully calibrated noise in data or model update. This paper focuses on some forms of DP mechanisms to use in FL, which include Gaussian, Laplacian, and Randomized Response mechanisms. It divides DP in FL into CDP, LDP, and Hybrid DP performance to ensure balance privacy and model utility. This work states the details of all these mechanisms are provided and issues such as privacy-utility trade-off, communication overhead, and scalability discussions are made.