Thermo-FL Enhances Edge LLM Fine-Tuning with Thermal-Aware Robustness.

This research introduces Thermo-FL, a novel framework designed to improve the fine-tuning of large language models (LLMs) on edge devices. It tackles two critical challenges: managing device thermal constraints and ensuring robustness against adversarial attacks during federated learning. The system dynamically adapts local training parameters, such as the active LoRA-layer fraction and update density, based on the device's temperature to prevent overheating and performance degradation. On the server side, Thermo-FL incorporates TERRA, a robust aggregation pipeline specifically designed for sparse LoRA updates. TERRA employs multiple techniques, including norm filtering, mask-aware directional validation, and adaptive clipping, to protect the global model from corrupted updates. Evaluations on both emulated and physical Jetson-based testbeds demonstrate Thermo-FL's ability to improve accuracy under adversarial conditions, stabilize device temperature, and reduce communication overhead while maintaining model utility.