Optimizing Federated Learning on TinyML Devices for Privacy Protection and Energy Efficiency in IoT Networks

Federated learning is presented as an effective solution to train artificial intelligence models on the Internet of Things networks without centralizing data, thus preserving privacy and minimizing security risks. However, its implementation in low-power devices, such as Tiny Machine Learning, faces significant challenges due to the processing, memory, and unstable connectivity limitations that characterize these environments. This study addresses these issues by developing a federated system optimized for Tiny Machine Learning devices, integrating differential privacy and encryption techniques adapted to their constraints. The methodology employed includes the evaluation of model precision and energy consumption in variable communication scenarios, as well as heterogeneous workloads. The results show that federated learning reduces energy consumption by 33% compared to the centralized approach, reaching an average of 100 mW. Furthermore, implementing differential privacy-maintained precision with a loss of only 1.2% demonstrates the feasibility of protecting sensitive data without substantially affecting performance. However, precision dropped by 8% in high communication loss scenarios, underlining the importance of stable connectivity for optimal system performance.