Alpha-FedAvg: Safeguarding Privacy and Enhancing Forensic Analysis in Federated Learning on Edge Devices

Abstract

In this paper, a novel federated learning algorithm for decentralized settings on edge devices— AlphaFedAvg—is introduced. Using an adaptive learning rate approach based on Lipschitz and Smoothness parameters, AlphaFedAvg dynamically modifies the learning rate for every node. Through federated averaging, the approach accomplishes model aggregation, exhibiting enhanced convergence and performance. An extensive test configuration includes using Kali Linux to simulate network assaults, an ESP32 microcontroller connected to a laptop equipped with a sound sensor, and Wireshark and Scapy for traffic analysis. The Alpha algorithm offers a privacypreserving solution by effectively identifying and thwarting network attacks without gaining access to user data. The algorithm's performance is demonstrated in a comprehensive report generated. Evaluation against IID and non-IID datasets, such as Edge-IIoTset, and comparison with other models validate AlphaFedAvg's efficacy in federated learning applications.