This paper presents DeepHardMark, a framework for embedding watermarks into Deep Neural Network (DNN) hardware accelerators. Unlike previous works that focus on protecting algorithmic intellectual property, this methodology defends the hardware itself. By embedding modifications into the accelerator's functional blocks, ownership can be verified using a secret key DNN and corresponding key sample. We utilize an $\ell_p$-box ADMM-based algorithm to co-optimize hardware overhead and algorithmic functionality. Evaluation on popular image classifiers demonstrates that DeepHardMark reliably embeds watermarks with minimal impact; for instance, a ResNet ImageNet classifier can be executed with an accuracy degradation of only 0.009%.
Introduction
The rapid adoption of Deep Neural Networks (DNNs) has led to a surge in specialized hardware accelerators, such as ASICs and FPGAs. However, the globalized semiconductor supply chain introduces significant risks, including IP piracy and unauthorized overproduction by third-party foundries. While software-level watermarking exists for DNN models, it fails to protect the underlying hardware. If a chip design is stolen, model-level protections are insufficient. DeepHardMark addresses this gap by "branding" the hardware logic itself, allowing owners to prove legitimacy without compromising chip efficiency or performance.
Methodology
DeepHardMark employs a "Trojan-inspired" approach to embed a permanent signature within the hardware. The process involves several key stages:
- Signature Design: A watermark is defined as a specific, incorrect output that is only triggered when the hardware processes a secret Key DNN and a secret Key Input.
- Hardware-Software Co-optimization: The framework treats watermarking as an optimization problem. Using an $\ell_p$-ADMM (Alternating Direction Method of Multipliers) framework, it identifies the optimal hardware functional blocks (e.g., Multiply-Accumulate or MAC units) for modification to minimize performance impact.
- Hardware Implementation: Small combinational logic circuits are embedded into the design. These circuits consist of a "trigger" to detect the key input and a "payload" to flip specific bits in the output, creating the verifiable signature.
A significant advantage of this approach is that it requires zero modifications to the model weights. The watermark resides entirely within the logic gates, ensuring the hardware remains general-purpose and can run other DNN models without activation.
Experimental Evaluation
The framework was validated using standard image classification benchmarks, including CIFAR-10, CIFAR-100, and ImageNet, across various architectures such as VGG, ResNet, and DenseNet.
Key results include:
- Effectiveness: An Embedding Success Rate (ESR) of 100% was achieved across all tests, ensuring reliable ownership verification.
- Minimal Performance Impact: Accuracy degradation was negligible. For a ResNet model on ImageNet, the accuracy drop was as low as 0.009%.
- Low Hardware Overhead:
- On FPGA implementations, logic overhead was approximately 0.18%, with a power increase of only 0.17%.
- ASIC simulations (using a "TinyTPU" design) showed a miniscule area overhead of 0.054%.
- High Fidelity: The "Fidelity Difference" (impact on non-key models) remained under 0.7% across all tests, proving the hardware's reliability for general use.
Conclusion
DeepHardMark provides a robust solution for protecting the intellectual property of DNN hardware. By leveraging hardware-aware optimization and stealthy logic modifications, it enables reliable ownership verification with near-zero overhead. Future work will explore extending this methodology to other hardware components beyond Matrix Multiply Units and evaluating its resilience against advanced Trojan-detection tools.
Summary updated