CASCADENCE: a layered cascade defense mechanism for federated learning

Abstract

Abstract This paper aims to enhance the security and robustness of Federated Learning (FL) systems through a multi-layered defense. We address the critical challenge of protecting distributed learning environments from adversarial attacks while maintaining high model performance during both the training and operational phases. The proposed framework is based on integrated approaches that utilize a Gaussian filter with Discrete Fourier Transform (DFT), adversarial training with differential privacy, JPEG compression, randomized smoothing, and adversarial logit pairing. It integrates multiple defense mechanisms based on system requirements, focusing on preserving model performance while ensuring robust protection during both training and testing phases. Our approach extends beyond existing solutions by introducing various staged defense implementations and analyzing their synergistic effects. Experimental results demonstrate that the proposed ensemble defense mechanism achieves the highest performance, maintaining 98.21% accuracy and an F1 score of 0.98 under attack conditions, compared to a baseline accuracy of 90.87%.