Research map / Vision-language and multimodal / Multimodal LLMs

Rescind: Countering Image Misconduct in Biomedical Publications with Vision-Language and State-Space Modeling

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds Rescind, a ~600K-image biomedical forgery benchmark generated with VLM-guided diffusion and LLaVA-Med verification, and Integscan, a state-space vision-language model for forgery localization.

Problem

Detecting and localizing manipulated images (duplication, splicing, removal) in biomedical publications, where existing forensic datasets are small, modality-limited, and use simplistic synthetic manipulations.

Approach

Generate semantically controlled biomedical forgeries by combining MedSAM masks, BiomedCLIP-derived prompts, and Stable Diffusion inpainting with a LLaVA-Med verification loop, and detect them with Integscan, a Mamba-style linear-attention SSM encoder with prompt-conditioned fusion (IntegBoost, Integration modules).

Key results

• Integscan reaches image-level MCC 0.776 on combined EDD and 0.793 on combined IDD on Rescind-T, above the best baselines (MONet 0.438, DF-PCT 0.569).
• On the Rescind-T EDD ablation, the full Integscan model attains F1 97.38 and IoU 94.96, above all ablated variants (IntegSSM replaced by CNN: F1 89.51).
• Rescind-V achieves the best image-quality scores among synthetic subsets (NIQE 3.31, PIQE 25.9, SSIM 0.894, LPIPS 0.116).

Surprising findings

• Models trained without FACS data perform poorly on FACS test samples, indicating modality-specific training data is needed.

Limitations

• Relies on VLM-inferred modality prompts that may misalign for ambiguous or low-quality inputs.
• Synthetic dataset may not capture compositional or multi-step real-world forgeries.
• Shared encoder across forgery types may limit task-specific specialization.

Experimental setup

Datasets: BioFors Rescind-P Rescind-T M3D-Synth

Benchmarks: Rescind-T BioFors

Metrics: MCC F1 IoU precision recall NIQE PIQE SSIM LPIPS

Baselines: DF-ZM DF-PCT DF-FMT BusterNet ManTra-Net TruFor MONet SIFT

Builds on: BioFors Mamba MedSAM BiomedCLIP LLaVA-Med Stable Diffusion inpainting

Closest papers by meaning

• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.85)
• FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.84)
• Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.84)
• OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL (ICML 2026 · similarity 0.84)
• ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.84)
• Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.84)