MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning

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

cs.AIMar 19, 2026

MedForge: Interpretable Medical Deepfake Detection via Forgery-aware Reasoning

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

Organizations: National University of Singapore · The Chinese University of Hong Kong · Hunan University · Xi’an Jiaotong University ·

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.

Figures & tables

Figure 1: Framework comparison. Left: computer vision detectors output binary decision, offering no clinically verifiable evidence. Right-bottom: post-hoc MLLM explainers (e.g., SIDA Huang et al. (2025b) ) may produce plausible-sounding but ungrounded rationales, including hallucinated visual details. Right-top: MedForge-Reasoner performs pre-hoc localized reasoning by first identifying suspicious regions (blue) and then generating medically coherent, visually verifiable rationales grounded in image evidence.

Figure 2: Overview of the MedForge-90K construction pipeline. The framework proceeds in three stages: medical image collection across three modalities, forgery generation via a Writer-Editor-Diagnoser loop, and expert-aligned annotation utilizing expert guidelines to generate hierarchical diagnostic reasoning.

Figure 3: MedForge-Reasoner Two-stage Training. SFT for cold-starting the reasoning format, followed by Forgery-aware GSPO. The GSPO stage introduces a reward function balancing visual grounding coverage and reasoning structure compliance to ensure the model localize the correct forgery region before reasoning.

Specialized Detectors
Methods	Real			Forgery: Implant			Forgery: Remove			Average
Methods	In-Domain	Cross-Forgery	Cross-Model	In-Domain	Cross-Forgery	Cross-Model	In-Domain	Cross-Forgery	Cross-Model	In-Domain	Cross-Forgery	Cross-Model
	Accuracy
SIDA-7B	$77.56_{(\uparrow 10.27)}$	$73.14_{(\uparrow 8.03)}$	$73.42_{(\uparrow 8.93)}$	$79.83_{(\uparrow 4.49)}$	$76.18_{(\uparrow 0.94)}$	$76.57_{(\uparrow 1.77)}$	$82.31_{(\uparrow 17.26)}$	$77.92_{(\uparrow 11.43)}$	$75.68_{(\uparrow 9.98)}$	$79.90_{(\uparrow 10.67)}$	$75.75_{(\uparrow 6.80)}$	$75.22_{(\uparrow 6.89)}$
SIDA-13B	$89.37_{(\uparrow 22.08)}$	$85.68_{(\uparrow 20.57)}$	$84.25_{(\uparrow 19.76)}$	$91.52_{(\uparrow 16.18)}$	$87.24_{(\uparrow 12.00)}$	$85.86_{(\uparrow 11.06)}$	$93.84_{(\uparrow 28.79)}$	$90.17_{(\uparrow 23.68)}$	$86.53_{(\uparrow 20.83)}$	$91.58_{(\uparrow 22.35)}$	$87.70_{(\uparrow 18.75)}$	$85.55_{(\uparrow 17.22)}$
FakeVLM	$86.48_{(\uparrow 19.19)}$	$81.86_{(\uparrow 16.75)}$	$76.94_{(\uparrow 12.45)}$	$88.93_{(\uparrow 13.59)}$	$83.57_{(\uparrow 8.33)}$	$79.82_{(\uparrow 5.02)}$	$91.27_{(\uparrow 26.22)}$	$86.35_{(\uparrow 19.86)}$	$82.51_{(\uparrow 16.81)}$	$88.89_{(\uparrow 19.66)}$	$83.93_{(\uparrow 14.98)}$	$79.76_{(\uparrow 11.43)}$

Table 1: Main Experiment - Forgery detection on MedForge-90K dataset. Methods are benchmarked against the Gemini 3 Pro, with arrows indicating performance differences (↑/↓) relative to it. Bold indicates the best result, and underline denotes the second-best.

MLLM-as-Judge	Gemini 3 Pro (%)				Qwen3-VL-Plus (%)
MLLM-as-Judge	LC	VH	MP	Avg.	LC	VH	MP	Avg.
SIDA-7B	34.6	27.9	54.3	38.9	71.3	53.4	78.5	67.7
SIDA-13B	40.2	32.6	61.6	44.8	75.6	57.7	82.1	71.8
FakeVLM	48.0	53.0	45.0	48.7	50.0	55.0	48.0	51.0
AIGI-Holmes	55.0	60.0	58.0	57.7	58.0	62.0	57.0	59.0
Qwen3-VL-Flash	68.2	57.3	85.9	70.5	87.6	73.9	91.0	84.2

Table 2: Evaluation of reasoning quality via MLLM-as-Judge. We report Logical Correctness (LC), Visual Hallucination (VH), Medical Professionalism (MP), and their Average score in percentage (%). Gray rows highlight the contribution of Forgery-aware GSPO.

(A) Contribution of Reasoning Components
Setting	Acc (%)	F1 (%)	IoU	Judge Score
Binary Classification	99.42	99.31	-	-
w/o Reasoning	99.31	99.10	0.30	-
w/o Bbox Grounding	99.31	98.97	-	53.9
Proposed	99.23	98.98	0.31	90.2
(B) Efficacy of Optimization Strategies

Table 3: Ablation studies on model components, optimization strategies, and backbone architectures.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Human Evaluation (%)
Method	LC	VH	MP	Avg.
SIDA-7B	52.9	52.3	53.4	52.9
FakeVLM	55.2	57.2	55.3	55.9
AIGI-Holmes	47.8	48.1	48.7	48.2
Qwen3-VL-Flash	75.7	75.8	75.6	75.7
Qwen3-VL-Plus	77.6	75.6	75.8	76.3

Appendix

Table 4: Human Evaluation on Reasoning Quality. We report Logical Correctness (LC), Visual Hallucination (VH), Medical Professionalism (MP), and their Average score in percentage (%). Evaluations were conducted by trained human annotators in a single-blind setting following the same protocol as the MLLM judges.

Method	Med-Diffu	Synth-Pneu	Synth-Bio	Med-Bana
Accuracy (%)
FakeVLM	50.0	50.0	50.0	54.1
AIGI-Holmes	51.6	51.2	49.8	49.6
Qwen3-VL-Plus	52.5	50.5	50.2	51.0
Gemini 3 Pro	90.5	80.7	74.9	70.8
Proposed	90.7	94.6	92.3	99.6

Appendix

Table 5: OOD Evaluation on Detection Performance. We assess generalizability across four unseen medical deepfake datasets. Bold indicates the best result, and underline denotes the second-best.

Modality	Implant	Removal	Area
CXR	99.37	99.47	6.05
MRI	98.65	98.63	4.29
Fundus	99.82	99.78	5.67

Appendix

Table 6: Failure-mode breakdown by modality and edit type. Values denote fake-image accuracy (%) for implant/removal and average ground truth forged-area ratio (%).

Rank	Modality	Pathology	Accuracy (%)	Errors
1	Fundus	Normal	97.60	55
2	Fundus	Myopia	98.68	1
3	MRI	Meningioma	98.73	7
4	MRI	Pituitary Adenoma	98.83	7
5	CXR	Pneumothorax	99.07	3
6	CXR	Pleural Effusion	99.10	3

Appendix

Table 7: Top-10 hardest authentic categories that are misclassified as deepfake (false positive).

Rank	Modality	Pathology	Accuracy (%)	Errors
1	MRI	Glioma	98.34	34
2	CXR	Fracture	98.39	10
3	MRI	Meningioma	98.68	29
4	CXR	Pneumothorax	98.87	7
5	MRI	Pituitary Adenoma	98.89	23
6	CXR	Atelectasis	99.10	6

Appendix

Table 8: Top-10 hardest fake categories that are misclassified as real (false negative).

Mean	Std	Median	95th	$>0.25$	$>0.5/0.75$
31.55	29.22	25.06	87.33	50.0%	27.2% / 11.7%

Appendix

Table 9: Localization IoU statistics on true-positive forgery test samples (

n=20{,}000

). IoU summary values are reported in percentage.

Deepfake Model	Dataset ID	Total	Implant	Remove
Stable Diffusion Inpainting (v1.5)	stable-diffusion-inpainting	6,303	3,239	3,064
Stable Diffusion XL Inpainting 0.1	stable-diffusion-xl-1.0-inpainting-0.1	6,380	3,273	3,107
FLUX.1-dev	flux.1-dev	6,399	3,273	3,126
Stable Diffusion 3.5 Large	stable-diffusion-3.5-large	6,383	3,273	3,110
Stable Diffusion 3.5 Medium	stable-diffusion-3.5-medium	6,383	3,273	3,110
GPT-Image	gpt	6,340	3,223	3,117

Appendix

Table 10: MedForge-90K image inventory. Per-editor RGB forgeries after the Writer–Editor–Diagnoser loop; Implant and Remove match the *-edit and *-remove splits. Deepfake models follow the presentation order in Appendix C.4 (inpainting

\rightarrow

advanced diffusion

\rightarrow

proprietary APIs), with Step1X-Edit listed last as an additional open instruction editor. Footer rows report forged aggregates, the authentic split total, and the full RGB inventory ( Implant / Remove are undefined for real and for the dataset-wide total).

Figure 4: Qualitative Forgery Explanation Comparison. Baselines fail due to severe hallucinations (SIDA citing “skin texture”) or missed diagnoses. While Gemini-3-Pro correctly detects the forgery using general visual clues, MedForge-Reasoner delivers superior clinically rigorous rationale , explicitly grounding the verdict in anatomical logic (e.g., “absence of mass effect”) rather than generic visual analysis.

Image Forgery Detection

Image Forgery Detection studies how to identify manipulated or synthetic images and locate the visual evidence of tampering.

See all 126 papers

Agentic Tool-Augmented Reasoning for Explainable Image Forgery Detection
Sep 30, 2026Zhiya Tan, Jing Huang, Changtao Miao +5Tool Use in VLMs Image Forgery Detection
CCDS Nanyang Technological University Singapore, Singapore · Ant Digital Technologies Ant Group Singapore, Singapore · Ant Digital Technologies Ant Group Hangzhou, China +3
HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes
Aug 5, 2026Orazio Pontorno, Luca Guarnera, Zahid Akhtar +1Explainable Deepfake Detection Medical Image Anomaly Detection
Department of Mathematics and Computer Science, University of Catania, Catania, Italy · Department of Electrical and Computer Engineering, State University of New York Polytechnic Institute, Utica, NY, USA
Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Models
Oct 13, 2026Sadif Ahmed, Soumya Swagata Biswas, Rifat Shahriyar +1Sycophancy in VLMs Medical VLMs
Bangladesh University of Engineering and Technology

Sep 30, 2026J/K move · Enter open · S save

Agentic Tool-Augmented Reasoning for Explainable Image Forgery Detection

Zhiya Tan, Jing Huang, Changtao Miao, Lin Tan+4

CCDS Nanyang Technological University Singapore, Singapore · Ant Digital Technologies Ant Group Singapore, Singapore · Ant Digital Technologies Ant Group Hangzhou, China+3

Tool Use in VLMs Image Forgery Detection Digital Image Forensics Image Forgery Localization

Open paper AskPDF arXiv

Conventional image forgery detection methods produce binary scores or pixel-level masks without interpretable evidence, while recent multimodal large language model (MLLM)-based approaches generate post-hoc explanations of predetermined classification results rather than reasoning from evidence. Inspired by the forensic workflow of human judicial experts, we propose Agentic Tool-Augmented Reasoning (ATAR), a framework integrating 22 specialized forensic tools across seven complementary domains to autonomously detect, localize, and explain image forgeries through multi-turn reasoning. A Dual-Stream Forensic Reasoning paradigm combines a high-level semantic anomaly path, which magnifies suspicious regions for fine-grained inspection, with a low-level forgery artifact path, which invokes forensic tools to extract objective evidence. We further introduce Forensics Curriculum Learning: during General Experience SFT, an automated teacher-student mentoring pipeline synthesizes multi-turn tool-usage reasoning trajectories; during Forensic Scene RL, a Tool Prior Curriculum guides early tool exploration and progressively transfers control to the agent, while a Structured Evidence Reward provides fine-grained process-level supervision. Experiments across IMDL, Deepfake detection, DMDL, and AIGC detection show that ATAR achieves 78.5% average image-level F1 on six zero-shot IMDL benchmarks, surpassing the strongest MLLM baseline by 11.8 percentage points, and remains competitive with specialized detectors on other tasks while producing substantially more faithful and grounded explanations.

Abstract

Image Forgery Detection

Explore similar work

Agentic Tool-Augmented Reasoning for Explainable Image Forgery Detection

HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes

Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Models