Edge devices now support advanced ML models, such as deep neural networks (DNNs), to process sensitive data locally, making model IP protection essential. This proposal introduces a hardware-based watermarking approach to safeguard TinyML model IP on resource-limited IoT devices. Unlike traditional methods that modify training data or outputs, our approach embeds watermarks within the model’s power side-channel signature by adding low-impact, redundant layers post-training. Key goals include designing power-trace-based encoding schemes, optimizing layer insertion with minimal performance impact, and evaluating reliability, transferability, and overhead on various IoT devices. This work offers an open-source solution to protect TinyML IP across edge platforms.