Fooling the Watchers: Breaking AIGC Detectors via Semantic Prompt Attacks

Run Hao, Peng YingarXiv 2025 · DOI · arXiv

Introduction

As text-to-image (T2I) models like Midjourney and Stable Diffusion reach unprecedented levels of realism, the boundary between human-captured and AI-generated content (AIGC) has blurred. This has necessitated the development of AIGC detectors designed to identify subtle mathematical "fingerprints"—artifacts in noise, color, or texture—left behind by generative algorithms. However, these detectors are not infallible.

Researchers have recently demonstrated an automated framework that systematically discovers "adversarial prompts." By shifting the focus from pixel-level noise to high-level semantic instructions, they have successfully bypassed both commercial and open-source detectors.

Overview of the AIGC detection bypass pipeline, showing the flow from prompt to detector.

The Blind Spots of AIGC Detection

Current detection methods often rely on the assumption that attackers will not actively try to mask generative artifacts. Most detectors are trained on static datasets and look for specific frequency distributions or texture patterns. Manual "prompt engineering" to find the weaknesses in these models is a slow, trial-and-error process.

The core problem addressed in this research is the lack of a systematic way to explore the "semantic prompt space." By identifying specific combinations of words that produce images appearing "real" to a detector, attackers can effectively hide the AI's signature without degrading the visual quality of the image for human viewers.

Methodology: Automated Semantic Attacks

The proposed framework automates the search for adversarial prompts using two primary components:

1. Grammar Tree Structure

To avoid random word guessing, the system organizes prompts into a logical hierarchy. It categorizes instructions into Portrait Attributes (e.g., age, gender), Stylistic Attributes (e.g., lighting, texture), and Miscellaneous Constraints. This structured approach allows the system to explore combinations that are both semantically coherent and likely to trigger detector blind spots.

The hierarchical grammar tree used to organize and combine prompt segments for systematic exploration.

Navigating the vast space of possible prompts requires an efficient search algorithm. The researchers employed a variation of Monte Carlo Tree Search (MCTS) called UCT-Rand. Unlike standard search algorithms that might converge too quickly on a single effective prompt, UCT-Rand uses weighted random sampling to ensure a diverse exploration of the semantic "attack" space.

Key Insights and Strategies

The research highlights several critical findings regarding how AIGC detectors can be fooled:

  • Semantics over Pixels: Traditional adversarial attacks involve changing individual pixels. This paper proves that semantic-level perturbations (changing words in the prompt) are just as effective and much harder to defend against because the resulting images look perfectly natural to humans.
  • Texture Injection: By prompting the model to include high-frequency textual patterns (e.g., "the rest of the page filled with clear text"), the attacker can confuse detectors like PatchCraft. These detectors often rely on analyzing the distribution of "rich" versus "poor" texture regions; the injected text "drowns out" the AI's subtle noise signatures.
  • Lighting and Blur Attacks: Instructions such as "dazzle" or "blur" alter the pixel intensity distribution. These effects mimic natural camera artifacts, which often mask the mathematical fingerprints that detectors are trained to find.

Results and Competition Success

The effectiveness of this semantic approach was validated in the Tencent AIGC Adversarial Detection Competition (2025), where the method secured first place. In this "black-box" scenario, the attackers had no knowledge of the underlying detector's architecture.

Key performance metrics included:

  • Commercial Bypass: Against the Zhuque detector, the system generated a high volume of images that consistently stayed below the 50% AI-detection threshold.
  • Open-Source Robustness: Against the PatchCraft detector, the flux-dev model achieved a 51% success rate within the first 100 images, with performance improving as search rounds increased.

Conclusion and Future Outlook

This research demonstrates that semantic-level perturbations are a potent threat to current AIGC detection systems. Because the resulting images look natural to humans, these attacks are significantly harder to defend against than traditional pixel-level adversarial noise.

However, the authors view this work as a stepping stone for defense. By using these automated tools to generate high-quality adversarial datasets, developers can train more robust, next-generation detectors capable of seeing through even the most sophisticated semantic masks.

Summary updated

Abstract

The rise of text-to-image (T2I) models has enabled the synthesis of photorealistic human portraits, raising serious concerns about identity misuse and the robustness of AIGC detectors. In this work, we propose an automated adversarial prompt generation framework that leverages a grammar tree structure and a variant of the Monte Carlo tree search algorithm to systematically explore the semantic prompt space. Our method generates diverse, controllable prompts that consistently evade both open-source and commercial AIGC detectors. Extensive experiments across multiple T2I models validate its effectiveness, and the approach ranked first in a real-world adversarial AIGC detection competition. Beyond attack scenarios, our method can also be used to construct high-quality adversarial datasets, providing valuable resources for training and evaluating more robust AIGC detection and defense systems.
BibTeX
@misc{hao2025fooling,
  author = {Run Hao and Peng Ying},
  title = {{Fooling the Watchers: Breaking AIGC Detectors via Semantic Prompt Attacks}},
  year = {2025},
  eprint = {2505.23192},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.23192},
}
Figure