Evaluating Differential Privacy in Federated Continual Learning: A Catastrophic Forgetting-Performance Tradeoff Analysis

Federated Learning (FL) enables model training directly on edge devices, thereby enhancing privacy by keeping data local. Despite advances in FL, the dynamic nature of data requires models to continuously adapt without forgetting previously acquired knowledge, thus leading to the adoption of Federated Continual Learning (FCL). However, this prior knowledge must be persistently stored in model parameters to mitigate catastrophic forgetting, which increases the risk of data or sensitive information exposure. Therefore, to bolster the privacy of local information, Differential Privacy (DP) is implemented as a privacy protection technique in FL, and its effectiveness is evaluated. Using the MNIST dataset, we assess the impact of noise on model accuracy under both Independent and Identically Distributed (IID) and Non-Independent and Identically Distributed (Non-IID) data distributions. We demonstrate that the proposed method not only adapts to changes in the operational environment but also significantly mitigates the risk of catastrophic forgetting.