Decentralized Federated Learning with Differential Privacy for Sensor Networks

Abstract

Decentralized Federated Learning (DFL) offers a promising approach to machine learning in resource-constrained environments like sensor networks, where data resides locally and communication is limited. However, traditional federated learning paradigms often fall short in providing robust privacy guarantees and are susceptible to attacks, particularly in the decentralized nature of sensor networks. This paper proposes a novel DFL framework that integrates differential privacy mechanisms at each sensor node alongside a Byzantine fault tolerance protocol. This combination ensures both privacy preservation and model integrity, addressing critical vulnerabilities inherent in existing decentralized learning systems. The framework utilizes a distributed aggregation strategy, minimizing communication overhead and enhancing resilience against malicious actors. The key contribution lies in the synergistic combination of differential privacy and Byzantine fault tolerance, providing a practical and secure solution for training machine learning models in decentralized sensor networks. The theoretical analysis demonstrates the privacy budget consumption and the effectiveness of the proposed protocol in mitigating the impact of Byzantine attacks. The system design prioritizes scalability and adaptability, crucial factors for deployment in diverse sensor network scenarios. This work provides a foundational approach for secure and efficient DFL in sensor networks, paving the way for innovative applications in areas such as environmental monitoring, smart cities, and industrial IoT.