BioTamperNet: Affinity-Guided State-Space Model Detecting Tampered Biomedical Images
Soumyaroop Nandi, Prem Natarajan
TL;DR
BioTamperNet is a new framework that uses affinity-guided attention mechanisms inspired by State Space Models to accurately detect duplicated and tampered regions in biomedical images, addressing the limitations of models trained on natural images.
Contribution
It introduces an affinity-guided self-attention and cross-attention modules based on linear attention mechanisms, enabling efficient and precise localization of tampered regions in biomedical images.
Findings
Significant improvement over baselines in detecting duplicated regions
Effective modeling of intra-image and cross-image similarities
Efficient linear attention mechanisms enable fine-grained localization
Abstract
We propose BioTamperNet, a novel framework for detecting duplicated regions in tampered biomedical images, leveraging affinity-guided attention inspired by State Space Model (SSM) approximations. Existing forensic models, primarily trained on natural images, often underperform on biomedical data where subtle manipulations can compromise experimental validity. To address this, BioTamperNet introduces an affinity-guided self-attention module to capture intra-image similarities and an affinity-guided cross-attention module to model cross-image correspondences. Our design integrates lightweight SSM-inspired linear attention mechanisms to enable efficient, fine-grained localization. Trained end-to-end, BioTamperNet simultaneously identifies tampered regions and their source counterparts. Extensive experiments on the benchmark bio-forensic datasets demonstrate significant improvements over…
Peer Reviews
Decision·ICLR 2026 Poster
The main strength is that the paper proposes a unified framework that simultaneously handles EDD, IDD, and CSTD in biomedical images by converting single-image tasks into pseudo image pairs, enabling one model to support all three forgery detection tasks without architectural changes.
1) The paper has poor writing and organization. The training dataset description should move from Section 2 to the experimental setup. 2) State Space Models are basic model to the proposed method. Section 2 should be renamed “Preliminaries” and include SSMs. 3) The paper claims BioTamperNet uses affinity-guided SSM attention modules but gives no clear explanation of how they integrate with or adapt SSMs. 4) Tables list baseline methods without citations which hurts readability and reproduc
The motivation is good, as the authors correctly identify the potentially severe long-term consequences of tampering in the biomedical domain, which could be as, if not more, damaging than in the natural image domain. The paper presents a commendable application and includes extensive comparisons against a range of existing methods.
Despite its promising direction, the manuscript suffers from several major flaws in its current form that significantly weaken its conclusions. These concerns relate to the validity of the experimental evaluation, the rigor of the theoretical claims, the completeness of the literature review, and the clarity of the methodology and figures. 1. The evaluation protocol: The authors state that the BioFors dataset provides annotations in the form of bounding boxes, not precise pixel-level masks. Ho
This work presents a novel and unified framework for tampering detection in biomedical images. The key innovation lies in its integration of affinity-guided attention with state-space models, yielding a computationally efficient and high-performing architecture. Extensive experiments across diverse biomedical modalities and benchmarks demonstrate clear superiority over prior works
1、The affinity-guided attention mechanism is conceptually appealing, but lacks qualitative visualizations (e.g., attention heatmaps, affinity evolution) that illustrate how the model distinguishes genuine vs tampered regions. 2、The model’s training relies heavily on synthetic duplications and GAN-generated patches. This may cause domain shift when applied to real, unannotated biomedical images. 3、Some equations and architectural diagrams are dense, reducing accessibility for readers. Section 3
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDigital Media Forensic Detection · Cell Image Analysis Techniques · Generative Adversarial Networks and Image Synthesis