BioTamperNet: Biomedical Tamper Detection

• BioTamperNet is a dual-framework system for detecting tampering in biomedical images and genomic networks using affinity-guided state-space models and cryptographic hash pipelines.
• It employs a ViT-based backbone with selective-scan SSMs and affinity-guided attention to localize duplicated regions in images while using centrality measures for network integrity.
• The approach achieves high accuracy and robustness—with superior MCC/F1 scores and minimal performance degradation under attacks—ensuring efficient, domain-specific bio-forensic defense.

BioTamperNet encompasses two key families of tamper-detection systems for biomedical data: (1) affinity-guided state-space models for localizing duplicated regions in biomedical images, and (2) hash-based tamper-evidence for complex genomic network graphs. Both paradigms address the acute challenge of detecting subtle, potentially devastating manipulations in biomedical digital artifacts, whether images or relational genomic data. This dual lineage is reflected in the named systems of (Nandi et al., 1 Feb 2026) (affinity-guided SSM models for images) and (Batool et al., 2017) (cryptographic-structural schemes for networks).

1. BioTamperNet for Biomedical Image Tamper Detection

BioTamperNet as introduced by (Nandi et al., 1 Feb 2026) is a Siamese, “feature-to-mask” deep network targeting detection and fine-grained localization of duplicated regions in tampered biomedical images, such as repeated cell clusters or gel bands. Contrary to generic image forensic models—largely trained on natural images—BioTamperNet leverages biomedical-specific data distributions and affine relationships. Its pipeline comprises:

• Feature extraction: A shared ViT-Base backbone pretrained on microscopy, blot/gel, FACS, and macroscopy modalities, generating patch-wise embeddings $V_1, V_2 \in \mathbb{R}^{B \times N \times C}$ ($N=H \cdot W$, $C=384$).
• Affinity-guided attention modules: Two distinct streams—affinity-guided self-attention (intra-image, capturing patch similarity within a single view) and affinity-guided cross-attention (inter-image, aligning duplicated regions between candidate pairs).
• Decoding: Lightweight convolutional decode heads map the enriched features to binary tamper masks $O_1, O_2 \in \mathbb{R}^{H \times W}$.
• SSM-inspired linear attention: Both attention streams are built atop selective-scan state-space model (SSM) blocks, providing efficient $O(N)$ sequence modeling, crucial for tractable large-patch grids.

By modeling local and cross-image patch affinities explicitly using SSM-derived attention, BioTamperNet addresses domain-specific image duplication artifacts that evade natural-image forensic detectors (e.g., ManTraNet, TruFor, SparseViT).

2. Architectural Components and Mathematical Formulation

2.1. Affinity Matrix Computation

Let $Q, K \in \mathbb{R}^{B \times N \times d}$ denote SSM-encoded query/key projections. Affinity matrices are computed via scaled dot-product attention: $$A_{\text{self}}(i,j) = \frac{\exp(Q_i \cdot K_j)}{\sum_{j'} \exp(Q_i \cdot K_{j'})}, \quad A_{\text{cross}}(i,j) = \frac{\exp(Q_i^{(1)} \cdot K_j^{(2)})}{\sum_{j'} \exp(Q_i^{(1)} \cdot K_{j'}^{(2)})}$$ Pre-attention features undergo selective-scan SSM transformation, where in discrete form: $$h_k = \bar{A} h_{k-1} + \bar{B} v_k, \quad y_k = \bar{C} h_k + \bar{D} v_k$$ Feature normalization is performed using rotary positional embeddings (RoPE) and ELU activations. The resulting "Similarity State-Space" output is used to construct affinity matrices $Aff_k = \bar{C}_k \bar{B}_k^\top$.

A spatial suppression kernel,

$$K(i,j,i',j') = \frac{(i-i')^2 + (j-j')^2}{(i-i')^2 + (j-j')^2 + \sigma^2}$$

is applied element-wise to $Aff_k$ to suppress trivial self-matches.

2.2. AGSSM and Cross-Attention Pipelines

Affinity-guided SSM-linear attention (AGSSM) modulates linear attention streams using affinity heatmaps. Each branch's outputs are averaged and projected for final self-attention enhancement: $$\tilde{V}_k = V_k + \text{AGSSM\_Self\_Attn}_k^{out}$$ Subsequent multi-head cross-attention exchanges information between $\tilde{V}_1$ and $\tilde{V}_2$, yielding the final cross-updated features. Theoretical analysis proves that the cross-attention localizes true duplicates with high confidence under a large enough affinity margin.

3. State-Space Model (SSM) Mechanisms

General SSMs are defined as

$h'(t) = A h(t) + B x(t), \quad y(t) = C h(t)$

Zero-order-hold discretization and Taylor expansion yield an efficient $O(N)$ kernel convolution implementation.

$$y_k = \bar{C} \sum_{n=1}^N \bar{B}_n \cdot x_{k-n} \iff y = k * x$$

This SSM-based linear attention approach allows large patch grids (e.g., $40 \times 40$) to be modeled tractably, constituting less than $4\%$ of total computation.

4. Training Regimen and Loss Composition

BioTamperNet is optimized end-to-end using a triple binary cross-entropy loss,

$$\mathcal{L} = w_{\text{self}} L_{\text{BCE}}^{\text{self}} + w_{\text{cross}} L_{\text{BCE}}^{\text{cross}} + w_{\text{fused}} L_{\text{BCE}}^{\text{fused}}$$

where $w$'s are set to 1.

Training utilizes:

• Synthetic data generated from pristine BioFors images (patch copy/inpainting, geometric/photometric/GAN perturbations)
• Input resolution $224 \times 224$
• AdamW optimizer: $lr=10^{-4}$, weight decay $0.05$, cosine decay schedule
• Pretraining (74 epochs), fine-tuning (100 epochs) on real images

IDD and CSTD modalities are handled via pseudo-pair augmentation strategies.

5. Performance Evaluation and Robustness

5.1. Datasets and Baselines

Benchmarks include BioFors (test split: 17,269 images; four modalities), RSIID microscopy, and Western-blot datasets. Evaluation metrics are pixel/image-level Matthews Correlation Coefficient (MCC) and F1-score.

Competing baselines: SIFT, ORB, BRIEF; dense matching (DF–ZM, DF–PCT, DF–FMT); deep nets (BusterNet, ManTraNet, TruFor, MONet, SparseViT, URN).

5.2. Results

Method	Microscopy Image/Pixel MCC	Blot/Gel	Macroscopy	FACS	Combined
SIFT	0.180 / 0.146	0.113 / 0.148	...	...	0.142 / 0.132
ORB	0.319 / 0.342	0.087 / 0.127	...	...	0.207 / 0.252
MONet	0.435 / 0.398	0.520 / 0.507	0.262 / 0.221	0.356 / 0.313	0.438 / 0.410
SparseViT	0.384 / 0.269	0.482 / 0.309	0.271 / 0.371	0.376 / 0.202	0.378 / 0.288
BioTamperNet	0.739 / 0.487	0.672 / 0.589	0.743 / 0.577	0.652 / 0.448	0.701 / 0.526

CSTD: F1=0.537/0.378, MCC=0.514/0.346 (BioTamperNet) vs. prior best MCC=0.173/0.092. On robustness benchmarks (RSIID, Western-Blot), BioTamperNet maintains pixel-level MCC superiority and demonstrates less than 5% MCC degradation under JPEG, contrast, and noise attacks—compared to 15–30% for other methods.

5.3. Ablation Studies

Ablating affinity, SSM, or attention modules reduces MCC by 10–20%. Adding a global SSM to affinity provides only marginal improvements, indicating that locality suffices for biomedical duplication detection.

6. Implementation, Complexity, and Optimizations

The system is implemented in PyTorch with selective-scan SSM modules realized as C++/CUDA kernels and employs rotary positional embeddings (RoPE) and depthwise-convolutional (DWConv) fused linear-attention. For input size $512 \times 512$, total parameter count is 36.7M with 29.6G FLOPs, providing a notably lower computational burden than TruFor (68.7M/236.5G) and SparseViT (50.3M/46.2G). Inference requires approximately 20 ms per image pair on a single NVIDIA V100.

7. BioTamperNet for Tamper-Evident Genomic Networks

An alternative instantiation of BioTamperNet (Batool et al., 2017) addresses tamper detection in complex genomic network graphs. Here, the primary concern is unauthorized topological changes—node/edge insertions, deletions, or attribute rewrites—that could jeopardize the veracity of relationship information in genomic repositories.

7.1. Core Methodology

• Representation: The network is modeled as $G=(V,E)$, where $V$ is the set of individuals (nodes) and $E$ is the set of biological/social relationships.
• Centrality-based feature extraction: Five centralities—degree ($C_D$), betweenness ($C_B$), closeness ($C_C$), eccentricity ($C_{\text{Ecc}}$), eigenvector ($C_{Ei}$)—are computed for each node.
• Hash pipeline: Each node's centralities are text-merged and SHA-1 hashed; the hashes are then concatenated and re-hashed to form a final 160-bit network digest ($H_{\text{network}}$).

$$M_v = \text{text-merge}\left(C_D(v),C_B(v),C_C(v),C_{\text{Ecc}}(v),C_{Ei}(v)\right)$$

$$h_v = H(M_v),\quad H_{\text{network}} = H(h_{v_1} \| h_{v_2} \| \ldots \| h_{v_n})$$

• Tamper evidence: Any unauthorized structural edit changes at least one node's centrality, altering $H_{\text{network}}$.

7.2. Empirical Assessment

Extensive tests on synthetic and canonical datasets (e.g., Zachary's Karate Club) yield a 100% detection rate for single-node/edge manipulations and zero false positives for legitimate updates. Computational overhead is $O(n+m)$ per check (for degree, closeness, etc.), rising for eigenvector computation in larger graphs.

7.3. Limitations and Future Directions

• The system's security is presently limited by reliance on SHA-1, which is now known to be weak; upgrading to SHA-256 or HMAC-hardened hashes is advised.
• It only authenticates graph topology, not rich node attributes (such as raw genomic sequences).
• Scaling to large, multiplex, or rapidly evolving graphs necessitates advances in centrality computation and update-efficient digesting.

8. Contextualization and Implications

BioTamperNet, across both image and graph-centric implementations, addresses a central need for rigorous, domain-tailored tamper detection in biomedical informatics. The image-centric mechanism unifies SSM-inspired affinity measures with end-to-end vision transformers to detect imperceptible region forgeries, achieving superior MCC/F1 accuracy and computational efficiency (Nandi et al., 1 Feb 2026). The network-digest paradigm realizes robust, topology-centric tamper evidence suitable for clinical genomic databases—at modest computational cost and with principled cryptographic integrity (Batool et al., 2017).

A plausible implication is that, as biomedical data manipulation grows more sophisticated and high-stakes, domain-specific, structurally aware defense schemes of the BioTamperNet vein will underpin next-generation bio-forensics.