MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning

Authors: Zhihui Chen, Kai He, Qingyuan Lei, Bin Pu, Jian Zhang, Yuling Xu, Mengling Feng

Published: 2026-03-19 07:38:11+00:00

AI Summary

This paper introduces MedForge, a data-and-method solution for interpretable medical deepfake detection, which tackles high-fidelity manipulations in medical scans. It establishes MedForge-90K, a large-scale benchmark of realistic lesion edits with expert-guided reasoning supervision. Building on this, MedForge-Reasoner performs pre-hoc localize-then-analyze reasoning, enhanced by Forgery-aware Group Sequence Policy Optimization (GSPO), to achieve state-of-the-art detection accuracy and trustworthy, expert-aligned explanations.

Abstract

Text-guided image editors can now manipulate authentic medical scans with high fidelity, enabling lesion implantation/removal that threatens clinical trust and safety. Existing defenses are inadequate for healthcare. Medical detectors are largely black-box, while MLLM-based explainers are typically post-hoc, lack medical expertise, and may hallucinate evidence on ambiguous cases. We present MedForge, a data-and-method solution for pre-hoc, evidence-grounded medical forgery detection. We introduce MedForge-90K, a large-scale benchmark of realistic lesion edits across 19 pathologies with expert-guided reasoning supervision via doctor inspection guidelines and gold edit locations. Building on it, MedForge-Reasoner performs localize-then-analyze reasoning, predicting suspicious regions before producing a verdict, and is further aligned with Forgery-aware GSPO to strengthen grounding and reduce hallucinations. Experiments demonstrate state-of-the-art detection accuracy and trustworthy, expert-aligned explanations.

Key findings

MedForge-Reasoner achieved state-of-the-art detection accuracy, outperforming strong baselines by over 7.65% for in-domain data and demonstrating significant robustness in out-of-distribution scenarios. The approach also delivered superior reasoning quality, reducing visual hallucinations by 16.2% and providing clinically rigorous, expert-aligned explanations for medical deepfake detection. The Forgery-aware GSPO module was crucial for improving explanation quality and reducing hallucinations.

Approach

The authors developed MedForge, comprising a specialized dataset and a novel MLLM-based detection method. They created MedForge-90K, a large-scale benchmark of medical image forgeries (lesion implant/removal) across 19 pathologies, annotated with expert-guided reasoning and gold edit locations. MedForge-Reasoner, an MLLM, is trained with a 'localize-then-analyze' objective using Supervised Fine-tuning (SFT) and further optimized by Forgery-aware GSPO to enforce visual grounding and reduce hallucinations in its explanations.

Datasets

MedForge-90K (custom benchmark built from MIMIC, ODIR, MultiEYE, Yale-Brain, and Brain-MRI datasets).

Model(s)

MedForge-Reasoner, based on Qwen3-VL-8B-Instruct.

Author countries

Singapore, China

← Previous