DeepTracer: Tracing Stolen Model via Deep Coupled Watermarks

DeepTracer represents a significant advancement in model watermarking techniques, introduced to combat the increasing threat of model stealing attacks. Traditional watermarking methods often fail to protect intellectual property, allowing adversaries to easily remove or bypass watermarks. DeepTracer's innovative approach involves a high-coupling model that integrates watermark tasks with primary tasks, ensuring that adversaries inadvertently learn the watermark while attempting to replicate the model's functionality. Additionally, the framework includes a sophisticated watermark samples filtering mechanism, enhancing the reliability of ownership verification. Extensive experiments demonstrate that DeepTracer surpasses existing methods in both effectiveness and robustness, marking a new state-of-the-art in the field. This development is crucial for safeguarding AI innovations and ensuring that creators can maintain ownership of their work.