Resource Efficient Deep Learning Hardware Watermarks with Signature Alignment

Joseph Clements, Yingjie LaoAAAI 2024 · OpenReview

Abstract

Deep learning intellectual properties (IPs) are high-value assets that are frequently susceptible to theft. This vulnerability has led to significant interest in defending the field's intellectual properties from theft. Recently, watermarking techniques have been extended to protect deep learning hardware from piracy. These techniques embed modifications that change the hardware's behavior when activated. In this work, we propose the first method for embedding watermarks in deep learning hardware that incorporates the owner's key samples into the embedding methodology. This improves our watermarks' reliability and efficiency in identifying the hardware over those generated using randomly selected key samples. Our experimental results demonstrate that by considering the target key samples when generating the hardware modifications, we can significantly increase the embedding success rate while targeting fewer functional blocks, decreasing the required hardware overhead needed to defend it.

The Challenge of Hardware IP Protection

Deep learning models are expensive to train, making them high-value targets for IP theft. While "algorithmic watermarks" (embedding secrets into a model's weights) are common, the underlying hardware accelerators are also vulnerable to piracy through supply chain leaks or unauthorized cloning.

The hardware watermarking lifecycle, illustrating the path from design and fabrication to piracy and subsequent verification.

Existing hardware watermarking techniques, such as the DeepHardMark framework, inject "backdoors" into silicon. When the hardware detects a specific "key" input, it produces an identifiable incorrect output to prove ownership. However, these prior methods relied on randomly selected key samples. This "blind" approach often required extensive, power-hungry hardware modifications to ensure the watermark triggered correctly, which could inadvertently degrade the performance of legitimate tasks.

Proposed Methodology: Signature Alignment

The authors introduce a Signature Alignment Algorithm that treats watermarking as a joint optimization problem between hardware and software. The system utilizes two primary logic circuits:

  • Activator: Monitors internal latent representations during computation to detect a specific "trigger" state.
  • Perturber: Once triggered, this circuit flips specific bits in the output to create the unique watermark signature.

Instead of using random images, the algorithm starts with a "seed" image and applies a tiny, invisible perturbation ($p$). Simultaneously, it selects the minimum number of functional blocks ($\eta$) to modify. By using $\ell_p$-ADMM (Alternating Direction Method of Multipliers), the authors find the absolute fewest hardware changes needed to reliably recognize the tweaked input. This "alignment" ensures the hardware "lock" and the input "key" fit perfectly with minimal overhead.

Experimental Evaluation

The methodology was evaluated using standard image classification benchmarks (CIFAR-10, CIFAR-100, ImageNet) across diverse architectures including ResNet, VGG, DenseNet, and Vision Transformers (ViT).

Key Results and Hardware Efficiency

The proposed method demonstrates a massive leap in efficiency compared to the state-of-the-art DeepHardMark.

Comparison of hardware overhead (Area, Cells, Power, Time) between the proposed Signature Alignment method and DeepHardMark.
  • Resource Efficiency: On the TinyTPU architecture, the new method reduced the required area and cell count by nearly 100x compared to previous benchmarks.
  • Zero-Impact Defense: Because modifications are sparse and specifically aligned, the impact on original model accuracy ($Acc_\Delta$) and fidelity ($Fid_\Delta$) is effectively 0%.
  • Power and Timing: Hardware overhead for power and propagation delay was reduced to nearly 0.00%, making the defense virtually "free" in terms of operating costs.
  • Success Rate: The method achieved an average Embedding Success Rate (ESR) of 92.6% even under tight hardware constraints.

Conclusion and Future Work

By co-optimizing the trigger input and hardware logic, signature alignment provides a robust, physically etched defense that is naturally resistant to software-level removal attacks. While currently requiring integration during the initial hardware design phase, future work may explore scaling this approach to non-vision tasks like Large Language Models (LLMs) and investigating resistance against adaptive adversaries who might attempt to reverse-engineer the alignment logic.

Summary updated

Abstract

Deep learning intellectual properties (IPs) are high-value assets that are frequently susceptible to theft. This vulnerability has led to significant interest in defending the field's intellectual properties from theft. Recently, watermarking techniques have been extended to protect deep learning hardware from privacy. These technique embed modifications that change the hardware's behavior when activated. In this work, we propose the first method for embedding watermarks in deep learning hardware that incorporates the owner's key samples into the embedding methodology. This improves our watermarks' reliability and efficiency in identifying the hardware over those generated using randomly selected key samples. Our experimental results demonstrate that by considering the target key samples when generating the hardware modifications, we can significantly increase the embedding success rate while targeting fewer functional blocks, decreasing the required hardware overhead needed to defend it.
BibTeX
@inproceedings{clements2024resource,
  author = {Joseph Clements and Yingjie Lao},
  title = {{Resource Efficient Deep Learning Hardware Watermarks with Signature Alignment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=yRwsOv1EYw},
}
Figure