Mitigating Gradient Inversion Attacks in Federated Learning with Frequency Transformation

Centralised machine learning approaches have raised concerns regarding the privacy of client data. To address this issue, privacy-preserving techniques such as Federated Learning (FL) have emerged, where only updated gradients are communicated instead of the raw client data. However, recent advances in security research have revealed vulnerabilities in this approach, demonstrating that gradients can be targeted and reconstructed, compromising the privacy of local instances. Such attacks, known as gradient inversion attacks, include techniques like deep leakage gradients (DLG). In this work, we explore the implications of gradient inversion attacks in FL and propose a novel defence mechanism, called Pruned Frequency-based Gradient Defence (pFGD), to mitigate these risks. Our defence strategy combines frequency transformation using techniques such as Discrete Cosine Transform (DCT) and employs pruning on the gradients to enhance privacy preservation. In this study, we perform a series of experiments on the MNIST dataset to evaluate the effectiveness of pFGD in defending against gradient inversion attacks. Our results clearly demonstrate the resilience and robustness of pFGD to gradient inversion attacks. The findings stress the need for strong privacy techniques to counter attacks and protect client data.