Abstract
Deep learning models represent significant intellectual property, yet they are vulnerable to theft via model extraction or unauthorized redistribution. Traditional black-box watermarking techniques often fail to balance stealth with robustness, resulting in triggers that are either easily detected or easily removed. This paper introduces ComMark, a novel framework that leverages frequency-domain compression (e.g., JPEG quantization) to create covert yet resilient ownership triggers. By shifting from pixel-level perturbations to global frequency signatures, ComMark ensures that watermarks survive common model-tampering attacks. Evaluations across image, audio, text, and video modalities demonstrate that ComMark achieves near-100% watermark success rates with minimal impact on primary task performance and high visual quality (PSNR > 40dB).
Introduction
The development of high-performance deep learning models requires substantial investment in data collection and computational power. Protecting this intellectual property is paramount. Model watermarking allows owners to embed a "trigger"—a specific input that elicits a secret, predictable output—to prove ownership through black-box API access. However, current methods suffer from a "stealth vs. strength" dilemma: overt triggers are robust but easily detected, while covert pixel-level triggers are fragile against common processing techniques.
Methodology
ComMark addresses the limitations of prior work by utilizing frequency-domain compression to generate triggers. This approach focuses on the global "behavior" of data rather than local pixel patterns.
Frequency-Domain Construction
The framework transforms inputs using the Discrete Cosine Transform (DCT) and discards high-frequency components through quantization, a process analogous to JPEG compression. The resulting "compressed" sample is visually indistinguishable from the original but contains a unique frequency signature that the model is trained to recognize.
Training and Robustness
To ensure the watermark is deeply integrated into the model's logic, ComMark employs a similarity loss (contrastive learning) to align watermark features with the target class in the latent space. Furthermore, attack simulation is integrated into the training pipeline, exposing the model to common preprocessing steps like cropping, rotation, and noise. This forces the model to maintain watermark sensitivity even under adversarial conditions.
Experimental Evaluation
The authors evaluated ComMark across a diverse range of modalities and datasets:
- Images: GTSRB, CIFAR-10/100, and VGGFace.
- Audio: Speech Commands and UrbanSound8K (using MP3-style compression).
- Text: IMDB and AG News, using linguistic "compression" (removing non-essential words).
- Video: UCF101 using H.264-style compression.
- Generative Tasks: Fashion MNIST and MSCOCO.
Key Results and Insights
ComMark demonstrates state-of-the-art performance in both effectiveness and robustness. It achieves a Watermark Success Rate (WSR) of nearly 100% across most benchmarks with a negligible impact (~1-2%) on primary task accuracy.
Robustness and Stealth
Unlike prior methods, ComMark is exceptionally resilient to model extraction and pruning. It maintains high WSR even when the model is pruned by 90%. Quantitatively, ComMark achieves a Peak Signal-to-Noise Ratio (PSNR) exceeding 40dB, making the watermarked samples virtually indistinguishable from clean data to the human eye. This high level of covertness ensures that the watermark remains hidden from adversaries while providing a strong proof of ownership.
Conclusion
ComMark represents a significant advancement in model IP protection. By leveraging the global properties of data compression, it provides a modality-agnostic solution that is both stealthy and robust. Future research may focus on adaptive compression strategies to further optimize the trade-off between watermark strength and visual imperceptibility.
Summary updated