MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning

ACL 2026 Main Conference

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

Paper (arXiv) Code, checkpoints & data (announcement)

Promotional summary

Our paper has been accepted to the ACL 2026 Main Conference: MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning.

As generative models become more powerful, medical deepfakes are becoming increasingly realistic: lesions can be injected into or removed from medical images while remaining visually plausible. This poses new risks to clinical safety, medical AI robustness, and the integrity of medical evidence. Existing methods mostly treat the task as binary real-vs-fake classification, but provide little insight into where the forgery is or why a sample is judged as manipulated.

To address this, we present MedForge, an interpretable framework for medical deepfake detection:

  • MedForge-90K: the first large-scale explainable medical deepfake dataset covering CT, MRI, and X-ray, with 19 lesion types and forgeries generated by 10 state-of-the-art deepfake models.
  • Each forged sample is annotated with expert-guided localization masks and clinical-grade explanations.
  • MedForge-Reasoner: an interpretable detector built on a Localize-then-Analyze chain-of-thought reasoning paradigm, trained with a vision-language model and Forgery-aware GSPO reinforcement learning.

Experiments show that MedForge-Reasoner achieves state-of-the-art detection performance while producing professional, localized, and verifiable medical explanations. Code, model checkpoints, and the full dataset will be released at this link.

A big thanks to the team and to Prof Mornin for his great support.

Figures

MedForge framework overview and comparison to prior medical deepfake detectors

Figure 1 — Overview of MedForge and comparison to specialized detectors and generic MLLM baselines.

MedForge-90K dataset construction pipeline

Figure 2 — MedForge-90K construction: real images, forgery generation, and reasoning annotation.

MedForge-Reasoner two-stage training with SFT and Forgery-aware GSPO

Figure 3 — MedForge-Reasoner two-stage training: reasoning cold-start (SFT) and Forgery-aware GSPO.

Qualitative comparison of forgery explanations across models

Figure 4 — Qualitative comparison of interpretable forgery judgments and rationales.

Main experimental results on MedForge-90K

Figure 5 — Main results: forgery detection on the MedForge-90K benchmark.