Security and Privacy‐Preserving Tiny Machine Learning

Abstract

This work examines the privacy and security issues with tiny machine learning (TinyML) and proposes potential solutions to improve dependability and trust in practical applications. A thorough analysis of current security risks is conducted, including data leakage, model extraction, and adversarial assaults. Furthermore, their suitability for TinyML environments is examined for privacy-preserving strategies, including secure enclaves, federated learning, and differential privacy. Because of their limited resources, TinyML systems present particular security and privacy issues that reduce the efficacy of conventional defences. New approaches to reducing these concerns, like decentralized learning, safe model deployment, and lightweight encryption, appear promising. This study offers a thorough examination of privacy and security concerns unique to TinyML, a comparatively unexplored subject within the larger field of artificial intelligence (AI) security. It advances secure edge AI research by bridging the gap between TinyML restrictions and contemporary security solutions.