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

Introduction

As Generative AI (GenAI) tools like Midjourney and ChatGPT become integrated into professional workflows, the creative industry faces a significant crisis regarding copyright and content provenance. While software-based watermarking and metadata tagging are common, they are easily stripped, modified, or bypassed by malicious actors. Researchers at the City University of Hong Kong have proposed a novel solution: RO-SVD, a hardware-level framework that anchors AI-generated content (AIGC) to the unique physical characteristics of the silicon on which it was created.

The Challenge: Device-Level Traceability

The fundamental problem in AIGC copyright is the lack of a permanent link between a digital asset and its source hardware. While Physically Unclonable Functions (PUFs) offer a potential solution by exploiting manufacturing variations in chips, they typically require rigid manufacturing constraints and complex error-correction logic to handle environmental noise. RO-SVD addresses these hurdles by utilizing reconfigurable Field-Programmable Gate Arrays (FPGAs) and a mathematical approach to isolate stable hardware identities without the need for heavy error-correction overhead.

The RO-SVD Framework

The RO-SVD architecture is a hardware-software co-design that transforms raw physical entropy into a robust security primitive through three primary stages:

1. Hardware Entropy Extraction

The system utilizes Ring Oscillators (ROs) implemented on an FPGA. Due to microscopic process variations during manufacturing, no two ROs vibrate at the exact same frequency, even when placed on identical chips. This variation creates a raw "entropy matrix" that serves as the device's unique physical signature.

2. Singular Value Decomposition (SVD) Processing

To make the raw hardware data usable, the framework applies Singular Value Decomposition (SVD). This mathematical process separates the entropy matrix into two distinct components:

  • Intrinsic Features (Largest Singular Values): These represent the stable, unique identity of the chip, functioning as a PUF for device authentication.
  • Stochastic Features (Smallest Singular Values): These represent unpredictable noise, which the framework repurposes as a True Random Number Generator (TRNG).

By using SVD, the researchers can mathematically "clean" the hardware signature, isolating the stable identity from environmental noise without the complex "Helper Data Algorithms" required by traditional PUFs.

3. Blockchain and Watermark Integration

The device identity is registered on a blockchain to create a permanent record. When the AI model generates content, the stochastic bits (TRNG) are used to generate a unique watermark, which is embedded into the image (e.g., via the Least Significant Bit). This transaction is logged on the blockchain, providing a verifiable path from the digital image back to the specific FPGA hardware.

Experimental Results

The researchers validated RO-SVD using a public dataset of 217 Xilinx FPGAs and real-world testing on PYNQ-Z2 development boards (Zynq XC7Z020 SoC). Key findings include:

  • Enhanced Randomness: Raw hardware responses often suffer from bias. RO-SVD improved the proportion of differences in generated random seeds from 2.96% to 43.8%, approaching the ideal 50% threshold for true randomness.
  • Uniqueness: The framework successfully distinguished between identical FPGA models from the same manufacturing batch, proving that SVD signatures are highly device-specific.
  • Performance and Efficiency: The hardware-accelerated SVD module handles 1024x1024 image matrices with millisecond-level latency. Furthermore, the security layer requires only 24,667 LUTs, making it a lightweight addition to modern AI-accelerated FPGAs.
  • Reconfigurability: Unlike fixed security chips, the FPGA-based approach allows the ROs to be moved to different "slices" of the chip. If a device's security is compromised, it can be reconfigured to generate a brand-new hardware fingerprint.

Conclusion and Future Work

RO-SVD provides a low-cost, robust, and reconfigurable mechanism for securing the AIGC pipeline. By moving the root of trust from software to hardware, it offers a more resilient defense against copyright infringement. While the current implementation focuses on image generation, future research will aim to optimize SVD component latency and expand the framework's application to video and 3D modeling, ensuring comprehensive protection for the next generation of digital media.

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 multidimensional 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. © 2024 IEEE.
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