<b>Evaluation of Privacy-Utility Trade-off in an Adaptive Differential Privacy-Infused Secure Aggregation (ADPSA) Model for Federated Learning in IoT Environments</b>

Abstract

A key challenge in privacy-preserving Federated Learning (FL) is the trade-off between achieving strong privacy guarantees and maintaining high model utility. While Differential Privacy (DP) offers a formal solution, static implementations often result in excessive noise, harming accuracy. This paper presents the comprehensive evaluation of the proposed Adaptive Differential Privacy-Infused Secure Aggregation (ADPSA) model, focusing on its ability to balance privacy and utility in IoT environments. The evaluation was conducted using the N-BaIoT dataset across 100 clients for 100 communication rounds. The model was assessed against four key metrics: final privacy budget (ε), resistance to gradient inversion (measured via PSNR and SSIM), membership inference risk (attack accuracy), and final model accuracy. The ADPSA model achieved a strong final privacy budget (ε = 4.17), excellent resistance to gradient inversion (PSNR = 9.4 dB, SSIM = 0.12), and near-optimal protection against membership inference (attack accuracy = 50.9%). It maintained a high final accuracy of 90.6%, with a performance degradation of just 1.98% compared to the non-DP baseline. The results demonstrate that the adaptive privacy mechanism successfully optimizes the privacy-utility trade-off, providing strong protection with minimal impact on model performance. These characteristics make ADPSA particularly suitable for resource-constrained IoT deployments where both privacy sensitivity and computational efficiency are critical.