Title: MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning

Abstract:

Text-driven image editing tools have recently achieved the capability to manipulate genuine medical scans with remarkable precision, allowing for the insertion or removal of lesions. This advancement poses significant risks to clinical trust and patient safety. Current defensive measures are insufficient for the healthcare sector; traditional medical detectors operate as opaque black boxes, while multimodal large language model (MLLM) explainers typically function as post-hoc systems. These explainers often lack specialized medical knowledge and are prone to hallucinating evidence, particularly when analyzing ambiguous cases.

To address these challenges, we introduce MedForge, a comprehensive solution encompassing both data and methodology for pre-hoc, evidence-based medical forgery detection. We present MedForge-90K, a large-scale benchmark featuring realistic lesion edits across 19 distinct pathologies. This dataset is supervised by expert-guided reasoning, incorporating doctor inspection guidelines and precise gold edit locations. Leveraging this benchmark, our MedForge-Reasoner employs a "localize-then-analyze" reasoning strategy, identifying suspicious regions prior to rendering a final verdict. Furthermore, the model is aligned with Forgery-aware GSPO to enhance grounding and minimize hallucinations. Experimental results indicate that our approach achieves state-of-the-art detection accuracy while providing trustworthy, expert-aligned explanations.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC