Q-Tag: Watermarking Quantum Circuit Generative Models

Yang Yang, Yuzhu Long, Han Fang, Zhaoyun Chen, Zhonghui Li, Weiming Zhang, Guoping Guo

TL;DR

As quantum circuit design shifts from manual engineering to automated generation via AI, protecting intellectual property (IP) becomes a moving target. Q-Tag is the first watermarking framework natively integrated into Quantum Circuit Generative Models (QCGMs), embedding ownership signals into the latent diffusion process to ensure that every generated circuit carries a robust, invisible, and verifiable digital signature.

What Problem Does This Paper Address?

Quantum computing resources are scarce, leading most researchers and companies to use "Quantum-as-a-Service" (QaaS) cloud platforms (like IBM Quantum or Amazon Braket). In this ecosystem, quantum circuits—the specific sequences of gates that perform a task—are high-value IP. However, once a circuit is uploaded to the cloud, it is vulnerable to unauthorized reuse or replication.

While "post-hoc" watermarking (adding extra gates to a finished circuit) exists, it is fragile. If a circuit is optimized or slightly edited, the watermark often breaks. More importantly, we are entering an era where Quantum Circuit Generative Models (QCGMs)—AI models that "dream up" optimal circuits—are becoming the standard. Protecting individual output circuits is no longer enough; we need to protect the generative process itself. The challenge is doing this without degrading the "fidelity" (the accuracy and performance) of the generated quantum circuits.

Methods

The authors introduce Q-Tag, a framework that embeds watermarks at the very beginning of the generative "birth" of a circuit. It relies on two novel components:

  1. Symmetric Sampling Mechanism (SSM): Most generative models (like Stable Diffusion) start with a block of random Gaussian noise. Q-Tag’s SSM partitions this noise into two symmetric regions ($P_0$ and $P_1$). To embed a binary watermark, the system "nudges" the initial noise samples into the region corresponding to the desired bit. Because the partitions are symmetric, the overall distribution remains Gaussian, meaning the generative model doesn't "notice" the change, and the quality of the output circuit remains high.
  2. Synchronization Restoration Mechanism (SRM): This is the defensive layer. If an attacker tries to "wash" the watermark by adding, deleting, or replacing gates, it causes "latent drift"—the internal representation of the circuit no longer aligns with the original watermark. SRM proactively uses zero-padding and a search strategy during the extraction phase to realign the data and recover the signature even from a tampered circuit.

The framework uses a Latent Diffusion Model (LDM) as its backbone, which is currently the state-of-the-art approach for generating complex data structures like quantum circuits.

Datasets

The researchers evaluated Q-Tag using a generative model trained on a large-scale quantum circuit dataset (following the architecture proposed by Furrutter et al. in Nature Machine Intelligence).

  • Evaluation Set: 1,000 independently generated watermarked circuits.
  • Attack Scenarios: The model was tested against four specific "structural attacks": gate replacement, gate appending, gate insertion, and gate deletion.
  • Baseline: They compared Q-Tag against existing structural watermarking methods (e.g., Roy et al.), which are applied after a circuit is created.

Key Insights and Ideas

  • Watermarking the "DNA," not the "Body": By embedding the watermark in the latent noise (the "DNA") before the diffusion process begins, the watermark becomes an intrinsic part of the circuit's structure rather than an appended "sticker."
  • Statistical Stealth: The "aha moment" here is the use of symmetric partitions. By ensuring the watermarked noise is statistically indistinguishable from pure random noise, the authors prove (both theoretically and via t-SNE visualization) that watermarking doesn't hurt the circuit's performance.
  • Resynchronization is the Real Battle: The paper recognizes that quantum circuits are "dynamic"—they are constantly being optimized by compilers. The SRM treats structural edits like a "desynchronization" problem, allowing the owner to find their watermark even if the "index" of the gates has shifted.

Key Results

  • High Fidelity: t-SNE plots show that watermarked and non-watermarked latents are perfectly overlapping, meaning the watermark is truly "invisible" to the model's performance metrics.
  • Robustness: Under "clean" conditions (no attacks), the True Positive Rate (TPR) for detection is 100%.
  • Resilience to Attacks: Even when circuits were subjected to aggressive structural modifications (like replacing or deleting gates), the watermark detection rate remained above 90%, and in some modification scenarios, reached 99.3%.
  • Superiority over Baselines: Traditional structural watermarking methods saw their detection rates plummet to ~20-40% under the same attacks, while Q-Tag remained stable.

Limitations and Future Work

While a major step forward, the authors note a few areas for improvement:

  • Handcrafted Heuristics: The current synchronization (SRM) relies on manual heuristics. Future versions could use machine learning to "learn" how to realign drifted latents more efficiently.
  • Capacity vs. Redundancy: There is a trade-off between how many bits you can hide (capacity) and how robust the watermark is. Pushing beyond 24 bits currently leads to a decline in robustness.
  • Trusted Extraction: The current model assumes the person extracting the watermark is a "trusted" entity with access to the original keys. Developing a "zero-knowledge" or public-verifiable extraction method would be the next logical step for decentralized quantum cloud security.

Summary updated

Abstract

Quantum cloud platforms have become the most widely adopted and mainstream approach for accessing quantum computing resources, due to the scarcity and operational complexity of quantum hardware. In this service-oriented paradigm, quantum circuits, which constitute high-value intellectual property, are exposed to risks of unauthorized access, reuse, and misuse. Digital watermarking has been explored as a promising mechanism for protecting quantum circuits by embedding ownership information for tracing and verification. However, driven by recent advances in generative artificial intelligence, the paradigm of quantum circuit design is shifting from individually and manually constructed circuits to automated synthesis based on quantum circuit generative models (QCGMs). In such generative settings, protecting only individual output circuits is insufficient, and existing post hoc, circuit-centric watermarking methods are not designed to integrate with the generative process, often failing to simultaneously ensure stealthiness, functional correctness, and robustness at scale. These limitations highlight the need for a new watermarking paradigm that is natively integrated with quantum circuit generative models. In this work, we present the first watermarking framework for QCGMs, which embeds ownership signals into the generation process while preserving circuit fidelity. We introduce a symmetric sampling strategy that aligns watermark encoding with the model's Gaussian prior, and a synchronization mechanism that counteracts adversarial watermark attack through latent drift correction. Empirical results confirm that our method achieves high-fidelity circuit generation and robust watermark detection across a range of perturbations, paving the way for scalable, secure copyright protection in AI-powered quantum design.
BibTeX
@misc{yang2026q,
  author = {Yang Yang and Yuzhu Long and Han Fang and Zhaoyun Chen and Zhonghui Li and Weiming Zhang and Guoping Guo},
  title = {{Q-Tag: Watermarking Quantum Circuit Generative Models}},
  year = {2026},
  eprint = {2602.23085},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.23085},
}
Figure