Enhancing Efficiency in Privacy-Preserving Federated Learning for Healthcare: Adaptive Gaussian Clipping With DFT Aggregator

Abstract

Machine learning’s exponential growth has transformed healthcare, with Federated Learning (FL) playing a pivotal role. Despite its significance, FL is vulnerable to privacy attacks. In response, researchers have integrated differential privacy (DP) into FL. Nevertheless, incorporating DP introduces challenges such as increased total communication costs and computational overheads due to the introduction of noise. This drawback renders FL with DP less viable for healthcare systems, characterized by numerous low-resource devices and network bandwidth constraints. To overcome this limitation, we propose integrating a Discrete Fourier Transform (DFT) aggregator post-noise addition to transform the gradient generated by local training before sending it to the central server. This process reduces the gradient size and provides rudimentary encryption. The evaluation results reveal the superior performance of our proposed method, demonstrating an enhanced accuracy ranging from 0.2% to 2% compared to existing differential privacy techniques, including RDP, DP-SGD, ZcDP, LDP-Fed, and DP-AdapClip. Our approach substantially reduces the total communication costs (ranging from 6% to 43% across different privacy budgets) with faster training times in healthcare datasets such as the PIMA Indian database and Breast Cancer Histopathology Images.

Keywords

• Federated learning
• differential privacy
• adaptive Gaussian clipping
• healthcare