Quantum-Inspired Fine-Tuning for Few-Shot AIGC Detection via Phase-Structured Reparameterization

Kaiyang Xing, Han Fang, Peng Wang, Zhonghui Li, Yang Claire Yang, Weiming Zhang, guoping guoarXiv 2026 · DOI · arXiv

In the escalating arms race between generative AI and detection tools, the "few-shot" regime is the ultimate battleground. How do you train a reliable detector when you only have a handful of examples of a new generative model's output?

A new paper, "Quantum-Inspired Fine-Tuning for Few-Shot AIGC Detection via Phase-Structured Reparameterization," suggests the answer might lie in the mathematical structures of quantum mechanics—even if you don't actually use a quantum computer.

TL;DR

Researchers have developed H-LoRA, a classical fine-tuning method inspired by quantum neural networks that improves AI-generated content (AIGC) detection by over 5% in low-data settings. By using the Hilbert Transform to inject "phase-aware" features into the standard LoRA adapter, they achieve the generalization benefits of quantum models without the massive computational overhead of quantum simulation.

What Problem Does This Paper Address?

Detecting AI-generated content (deepfakes) is difficult because generative models evolve rapidly. Traditional fine-tuning requires large datasets, but in the real world, we often need to adapt detectors to new models using only a few hundred samples (few-shot learning).

While Low-Rank Adaptation (LoRA) is the industry standard for efficient fine-tuning, it often struggles to generalize when data is scarce. Conversely, Quantum Neural Networks (QNNs) are known for excellent generalization in few-shot regimes due to their unique geometric properties. However, QNNs are incredibly slow to simulate on classical hardware, making them impractical for large-scale tasks like analyzing high-resolution images or long audio clips.

The authors set out to bridge this gap: Can we distill the "secret sauce" of quantum generalization into a fast, classical algorithm?

Methods

The researchers started by building Q-LoRA, which embeds a lightweight 4-qubit QNN directly into the bottleneck of a LoRA adapter. While Q-LoRA performed well, it was painfully slow (taking 30 minutes per epoch vs. seconds for standard LoRA).

By analyzing why Q-LoRA worked, they identified two key "inductive biases":

  1. Phase-Aware Representations: Quantum states naturally encode information in both amplitude and phase.
  2. Norm-Constrained Transformations: Quantum gates are unitary (norm-preserving), which stabilizes optimization.

To replicate this classically, they created H-LoRA. Instead of a quantum circuit, H-LoRA uses the Hilbert Transform—a staple of signal processing—to convert real-valued features into "analytic signals" composed of orthogonal amplitude and phase components. This allows the model to capture subtle structural patterns in AI-generated images and audio that standard linear layers might miss.

Datasets

The authors evaluated their methods across two distinct domains:

  • Image Detection: They used the AIGCDetectionBenchMark, training on Stable Diffusion v1.4 and testing generalization across a "who's who" of generators: Midjourney, Wukong, VQDM, Glide, ADM, and Stable Diffusion v1.5.
  • Audio Detection: They utilized the ASVspoof 2019 LA dataset to test few-shot audio forgery detection, using OpenAI’s Whisper as the backbone model.

Key Insights and Ideas

  • The "Quantum" Advantage is Structural: The paper argues that the benefit of QNNs isn't necessarily "quantum magic" but rather the specific way they structure information. By mimicking this structure (phase-encoding) classically, you get the same performance boost.
  • Phase Matters for Deepfakes: AI generators often leave subtle artifacts in the "phase" or frequency domain of an image or audio clip. H-LoRA explicitly forces the model to look at these components, making it harder for generated content to "blend in."
  • Stability through Constraints: By enforcing norm-constrained transformations, H-LoRA acts as a regularizer, preventing the model from overfitting to the small training set—a common pitfall in few-shot learning.

Key Results

  • Accuracy Boost: In the most challenging 200-sample regime, H-LoRA outperformed standard LoRA by an average of 5.63% accuracy across various image generators.
  • Massive Speedup: While Q-LoRA and H-LoRA achieved similar accuracy, H-LoRA was roughly 500x faster to train. For a 200-sample task, Q-LoRA took ~2088 seconds per epoch, while H-LoRA took only 4.07 seconds.
  • Cross-Domain Success: H-LoRA isn't just for images. In 50-shot audio detection, it achieved 89.33% accuracy, significantly higher than standard LoRA’s 80.34%.
  • Zero Extra Parameters: Unlike Q-LoRA, which adds a small number of trainable parameters, H-LoRA introduces zero additional trainable parameters over standard LoRA, making it a "drop-in" upgrade.

Limitations and Future Work

While H-LoRA is a significant step forward, the authors note a few areas for growth:

  • Backbone Dependency: The method was tested on CLIP (images) and Whisper (audio). Its effectiveness on even larger LLMs or different architectures remains to be seen.
  • Fixed Phase Shift: The current implementation uses a standard 90-degree phase shift via the Hilbert Transform. Future work could explore learnable phase-shifting operations to see if the model can discover even better representations.
  • Complexity: While faster than quantum simulation, the Hilbert Transform (often implemented via FFTs) still adds a small amount of computational complexity compared to a simple matrix multiplication.

In summary, H-LoRA proves that we don't need a quantum computer to benefit from quantum thinking. By borrowing the concept of phase-aware representations, we can build tougher, faster detectors for the next generation of AI content.

Summary updated

Abstract

Recent studies show that quantum neural networks (QNNs) generalize well in few-shot regimes. To extend this advantage to large-scale tasks, we propose Q-LoRA, a quantum-enhanced fine-tuning scheme that integrates lightweight QNNs into the low-rank adaptation (LoRA) adapter. Applied to AI-generated content (AIGC) detection, Q-LoRA consistently outperforms standard LoRA under few-shot settings. We analyze the source of this improvement and identify two possible structural inductive biases from QNNs: (i) phase-aware representations, which encode richer information across orthogonal amplitude-phase components, and (ii) norm-constrained transformations, which stabilize optimization via inherent orthogonality. However, Q-LoRA incurs non-trivial overhead due to quantum simulation. Motivated by our analysis, we further introduce H-LoRA, a fully classical variant that applies the Hilbert transform within the LoRA adapter to retain similar phase structure and constraints. Experiments on few-shot AIGC detection show that both Q-LoRA and H-LoRA outperform standard LoRA by over 5% accuracy, with H-LoRA achieving comparable accuracy at significantly lower cost in this task.
BibTeX
@misc{xing2026quantum,
  author = {Kaiyang Xing and Han Fang and Peng Wang and Zhonghui Li and Yang Claire Yang and Weiming Zhang and guoping guo},
  title = {{Quantum-Inspired Fine-Tuning for Few-Shot AIGC Detection via Phase-Structured Reparameterization}},
  year = {2026},
  eprint = {2603.02281},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.02281},
}
Figure