PRIVACY-PRESERVING FEDERATED LEARNING WITH DIFFERENTIAL PRIVACY AND EDGE AI FOR RESOURCE-CONSTRAINED IOT DEVICES

Abstract

As the Internet of Things (IoT) revolution has swept the world, the data produced is vast and distributed, bringing with it opportunities for smart data analysis and concerns about privacy and security. Conventional distributed and centralized machine learning systems involve sending sensitive data to cloud servers, which can lead to communication delays and data breaches. The problem has been addressed by offering a new approach capable of collaborative model training without sharing raw data, known as Federated Learning (FL). Federated learning is still susceptible to information leakage, poisoning attacks, and resource limitations of edge devices, however. This paper introduces a Privacy-Preserving Federated Learning (PPFL) framework for resource constrained IoT applications utilizing Differential Privacy (DP), Secure Aggregation and Edge AI. The proposed framework allows for distributed model training while maintaining privacy preserving model updates and secure aggregation mechanisms. The model convergence, privacy-accuracy trade-off, communication overhead, energy consumption, resource usage, model robustness, scalability and feasibility of deployment were comprehensively evaluated through experiments. The experimental results show that the proposed framework can guarantee stable convergence with different privacy budgets and provide a good trade-off between privacy preservation and predictive performance. Under relaxed privacy settings, the framework obtains the accuracy of up to 86% in models, while still assuring strong privacy guarantees and resistance to adversarial attacks. Moreover, the communication efficient mechanisms resulted in less overhead in the network, while the evaluations in the edge deployment demonstrated the implementation of the framework with resource constrained devices. The results show that the proposed PPFL framework can be a practical solution for privacy-sensitive IoT applications such as a smart healthcare system, a smart city system, and an industrial IoT system.