通过多模态大语言模型中的扎根推理实现可解释且可靠的 AI 生成图像检测
Interpretable and Reliable Detection of AI-Generated Images via Grounded Reasoning in MLLMs
AI总结：
针对 AI 生成图像检测缺乏可解释性的问题，本文构建带边界框与伪影描述的数据集，并通过多阶段微调 MLLMs，实现准确检测、视觉定位与文本解释的协同提升。
摘要：
图像生成技术的快速发展，使得对可解释且鲁棒的检测方法的需求日益迫切。尽管现有方法通常能够取得较高准确率，但它们往往像黑箱一样运行，无法提供人类可理解的依据。多模态大语言模型（MLLMs）虽然最初并非为伪造检测而设计，却展现出强大的分析与推理能力。经过适当微调后，它们可以有效识别 AI 生成图像，并给出有意义的解释。然而，现有 MLLMs 仍受幻觉问题困扰，且往往无法使其视觉解释与实际图像内容及人类推理保持一致。为弥合这一差距，我们构建了一个 AI 生成图像数据集，其中标注有边界框和突出合成伪影的描述性字幕，为符合人类推理的视觉-文本扎根推理奠定基础。随后，我们通过多阶段优化策略微调 MLLMs，逐步平衡准确检测、视觉定位和连贯文本解释等目标。所得模型在检测 AI 生成图像和定位视觉缺陷两方面均取得优异性能，显著优于基线方法。
英文摘要：
The rapid advancement of image generation technologies intensifies the demand for interpretable and robust detection methods. Although existing approaches often attain high accuracy, they typically operate as black boxes without providing human-understandable justifications. Multi-modal Large Language Models (MLLMs), while not originally intended for forgery detection, exhibit strong analytical and reasoning capabilities. When properly fine-tuned, they can effectively identify AI-generated images and offer meaningful explanations. However, existing MLLMs still struggle with hallucination and often fail to align their visual interpretations with actual image content and human reasoning. To bridge this gap, we construct a dataset of AI-generated images annotated with bounding boxes and descriptive captions that highlight synthesis artifacts, establishing a foundation for human-aligned visual-textual grounded reasoning. We then finetune MLLMs through a multi-stage optimization strategy that progressively balances the objectives of accurate detection, visual localization, and coherent textual explanation. The resulting model achieves superior performance in both detecting AI-generated images and localizing visual flaws, significantly outperforming baseline methods.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
