RO-SVD: A Reconfigurable Hardware Copyright Protection Framework for AIGC Applications

Zhuoheng Ran, Muhammad A. A. Abdelgawad, Zekai Zhang, Ray C. C. Cheung, Hong Yan

TL;DR

Researchers have developed RO-SVD, the first hardware-based framework designed to bake copyright traceability directly into AI-generated content (AIGC) at the device level. By combining the unique physical variations of FPGA chips with Singular Value Decomposition (SVD) and blockchain technology, they’ve created a system that generates unique, unclonable digital fingerprints for every AI-generated image.

What Problem Does This Paper Address?

As generative AI tools like Midjourney and Sora flood the internet with content, determining who (or what device) created a specific image has become a legal and ethical minefield. Current solutions, like metadata tracking, are easily stripped or modified.

The authors argue that for copyright to be robust, it needs to be tied to the hardware entropy—the microscopic, unintentional physical differences in a computer chip's circuitry that make every piece of silicon unique. However, using this "hardware noise" for AIGC is difficult because the raw data is often unstable and massive in volume. The challenge is to extract a reliable, unique identity from a noisy hardware source and link it efficiently to a massive stream of AI-generated data.

Methods

The framework, called RO-SVD, utilizes a hardware-software co-design on Field-Programmable Gate Arrays (FPGAs).

  1. Entropy Generation: The system uses groups of Ring Oscillators (ROs). Because of tiny manufacturing variations, these oscillators run at slightly different speeds on every chip. This creates a unique "silicon fingerprint."
  2. Singular Value Decomposition (SVD): This is the "secret sauce." The raw entropy data is treated as a matrix. The researchers use SVD to decompose this matrix into two parts:
  3. Intrinsic Features: The stable, unique signature of the chip (extracted via the largest singular values).
  4. Stochastic Features: The random noise (extracted via the smallest singular values).
  5. Dual-Purpose Output: The intrinsic features become a Device ID for authentication, while the stochastic features serve as a True Random Number Generator (TRNG) for creating unique transaction seeds.
  6. Blockchain Integration: These IDs and seeds are hashed and stored on a blockchain, creating an unalterable ledger of which device created which piece of content.
  7. Watermarking: The final ID is embedded into the AI-generated image using Least Significant Bit (LSB) watermarking, ensuring the "mark" is invisible but traceable.
Fig. 1. The overall flow of the proposed framework with the least significant bit (LSB) lossless application side.

Datasets

The study utilized a large-scale public dataset of frequency measurements from 217 Xilinx 28nm FPGAs (originally collected by the University of Belfast). This dataset includes 100 repeated measurements from over 6,500 ring oscillators per chip, allowing the authors to simulate and validate how their SVD approach handles die-to-die and intra-die variations.

Key Insights and Ideas

  • Hardware as a Low-Rank Matrix: The most significant insight is that hardware entropy responses exhibit approximate low-rank properties. By applying Truncated SVD (TSVD), the researchers could effectively "denoise" the hardware signature, separating the permanent physical identity from transient environmental noise (like temperature fluctuations).
  • Separation of Concerns: Instead of trying to make the hardware perfectly stable (which is hard and expensive), they use math (SVD) to separate the stable "identity" from the useful "randomness."
  • Reconfigurability: Because the system is built on FPGAs, the security primitives can be updated or "renewed" by changing the logic constraints, providing a flexible defense against potential exploits.
Fig. 3. Pattern recognition of hardware entropy sources.

Key Results

  • Improved Randomness: Without processing, the raw hardware response had a difference proportion of only 2.96% between rows. After applying the RO-SVD method, this increased to 43.80%, approaching the ideal 50% for high-quality random number generation.
  • NIST Compliance: The generated random seeds passed the NIST SP 800-22 statistical test suite, confirming they are sufficiently unpredictable for cryptographic use.
  • Hardware Efficiency: The prototype was implemented on PYNQ-Z2 boards. While SVD is computationally intensive, the authors optimized the design using HLS (High-Level Synthesis) and AXI protocols to minimize latency, achieving millisecond-level data exchange with DDR memory.
  • Power Consumption: The total power consumption for the SVD and matrix operations was measured at approximately 3.85W, making it suitable for integration into existing AI-accelerated edge devices.

Limitations and Future Work

The authors acknowledge that matrix decomposition is resource-intensive. While they successfully shifted much of the computational burden to the "traceability side" (where resources are more abundant), the initial generation on the device still requires significant FPGA logic (LUTs and Registers).

Future work will focus on:

  • Enhancing the computing efficiency of the SVD component to increase overall system speed.
  • Exploring how this framework can be applied to other forms of AIGC beyond 2D images, such as video or 3D models.
  • Testing the robustness of the LSB watermarking against common image attacks (compression, cropping, etc.) in "non-lossless" scenarios.

Summary updated

Abstract

The dramatic surge in the utilisation of generative artificial intelligence (GenAI) underscores the need for a secure and efficient mechanism to responsibly manage, use and disseminate multi-dimensional data generated by artificial intelligence (AI). In this paper, we propose a blockchain-based copyright traceability framework called ring oscillator-singular value decomposition (RO-SVD), which introduces decomposition computing to approximate low-rank matrices generated from hardware entropy sources and establishes an AI-generated content (AIGC) copyright traceability mechanism at the device level. By leveraging the parallelism and reconfigurability of field-programmable gate arrays (FPGAs), our framework can be easily constructed on existing AI-accelerated devices and provide a low-cost solution to emerging copyright issues of AIGC. We developed a hardware-software (HW/SW) co-design prototype based on comprehensive analysis and on-board experiments with multiple AI-applicable FPGAs. Using AI-generated images as a case study, our framework demonstrated effectiveness and emphasised customisation, unpredictability, efficiency, management and reconfigurability. To the best of our knowledge, this is the first practical hardware study discussing and implementing copyright traceability specifically for AI-generated content.
BibTeX
@inproceedings{ran2024ro,
  author = {Zhuoheng Ran and Muhammad A. A. Abdelgawad and Zekai Zhang and Ray C. C. Cheung and Hong Yan},
  title = {{RO-SVD: A Reconfigurable Hardware Copyright Protection Framework for AIGC Applications}},
  booktitle = {ASAP},
  year = {2024},
}
Figure