ALFadapt: Mitigating Catastrophic Forgetting in Federated Learning

Catastrophic forgetting, driven by concept drift, challenges federated learning (FL), especially in dynamic environments. Existing layer-freezing methods focus on communication efficiency but often overlook concept drift’s impact on individual neural network layers. We propose ALFadapt (Automatic and Adaptive Layer Freezing) to mitigate catastrophic forgetting by dynamically freezing stable layers and allowing trainable layers to adapt to evolving data. Experimental results show that ALFadapt significantly reduces accuracy loss during scenario transitions and improves model resilience when revisiting previous scenarios. This method offers a robust solution for environments characterized by concept drift and non-IID data distributions.