MULTI-LAYER PRIVACY FRAMEWORK FOR TINY FEDERATED LEARNING ON RESOURCE-CONSTRAINED DEVICES

Abstract

Ensuring data privacy in Federated Learning (FL) is a critical challenge, especially for resource-constrained devices like single-board computers and microcontrollers, which are widely used in edge computing and IoT networks. This article presents a novel multi-layer privacy framework designed for Tiny Federated Learning (TFL), combining homomorphic encryption and differential privacy to provide robust data protection while maintaining computational efficiency. Our approach addresses privacy challenges in anomaly-based network intrusion detection systems by securing local model updates and preventing data inference attacks without compromising performance. We optimize encryption algorithms to minimize energy consumption and communication overhead, making them suitable for deployment on low-power devices. The proposed framework is evaluated on a network of Raspberry Pi devices using real-world datasets, demonstrating significant improvements in data security and system efficiency compared to traditional privacy-preserving methods. Our results highlight the balance between maintaining model accuracy and ensuring privacy, paving the way for secure and scalable FL implementations in constrained environments.