PaperlayerView exact revision on arXiv

Source-linked AI summary

GAN-Blot: A Controllable Structure-Style Synthesis Benchmark for Western Blot Forensics

Hao-Chiang Shao, Fong-Yi Lin, Te-An Chien, TianYu Chen, Da-Jhong Chen

arXiv:2609.06619v1cs.CV

TL;DR

WB forensics lacks standardized appearance definitions, datasets, and controllable generators for evaluating AI-fabricated images. GAN-Blot separates band structure from style and builds a synthetic benchmark; its images show high fidelity yet can fool experts and are not reliably distinguished by existing tools.

  • Problem

    WB forensics lacks standardized appearance definitions, datasets, and controllable generation frameworks, limiting systematic evaluation of AI-generated WB images.

    Source: §I · pp. 1–2

  • Method

    GAN-Blot independently conditions WB synthesis on a protein-band mask and style reference, using dual-path training and style-alignment losses without predefined appearance attributes.

    Source: §I · p. 2, A. Overview · p. 3

  • Results

    GAN-Blot generates images with high fidelity in band structure and visual style that can fool experts under blind inspection and are not reliably distinguished by existing detectors and screening platforms.

    Source: §I · p. 2

  • Takeaways & Limitations

    The synthetic images provide challenging controlled cases for validating and developing WB forensic methods.

    Source: §I · pp. 1–2

Abstract

from arXiv · show

Western blot (WB) images are widely used as key evidence in biomedical research. Recent scientific misconduct cases reveal that WB imagery is increasingly fabricated, making WB forensics a major concern for research integrity. However, while the progress of forensic detection techniques often relies on advances in forgery-generation techniques, the development of WB forensic techniques has been hindered by the lack of standardized appearance attribute definitions, image datasets, and controllable generation frameworks for WB imagery. To address this limitation, we present a controllable WB image synthesis framework, named GAN-Blot, for generating realistic synthetic WB images. We introduce a formulation that decomposes a WB image into a structure component and a style-reference component, enabling independent control over local protein-band geometry and the global visual appearance of a synthetic WB image. GAN-Blot integrates a dual-path autoencoding design with several style-alignment loss terms to enable implicit control over structure-style synthesis without predefined semantic appearance attributes. We further contribute a synthetic WB dataset containing more than 46K images and propose four evaluation protocols for controllable WB synthesis. Extensive experiments show that GAN-Blot can generate WB images with high fidelity in both protein-band structure and visual style. Under blind inspection, the generated images can fool domain experts and are not reliably distinguished from authentic WB images by existing detectors and screening platforms. These results demonstrate their utility as challenging controlled cases for validating and developing WB forensic methods.

I. INTRODUCTION

Western blot images are consequential forensic targets, but methods developed for natural images and the lack of standardized controllable-generation benchmarks leave AI-generated WB forensics difficult to evaluate. GAN-Blot addresses this gap by separating band structure from visual style and supplying synthetic images and evaluation protocols.

I. INTRODUCTION

  • WB images are frequent falsification targets, while most AI-image detectors address natural or face images and WB methods mainly address conventional manipulations.The paper frames AI-generated WB images as a distinct forensic scenario requiring assessment of whether patterns could arise from authentic experiments.Source: §I · pp. 1–2
  • Standardized generation benchmarks and evaluation protocols remain lacking, while WB-specific attributes do not map readily to natural-image semantic controls.These attributes include bands, lane traces, background intensity, and gel or membrane artifacts.Source: §I · p. 2, Figure 2 · p. 2
  • GAN-Blot independently conditions synthesis on a band-contour mask and a style-reference image, controlling local band structure and global appearance without predefined semantic attributes.The stated controllable outputs include band structures, lane patterns, intensity distributions, and gel or membrane styles.Source: §I · p. 2
  • The paper contributes a synthetic WB benchmark and four evaluation protocols covering controllability, target-conditioned resynthesis, expert assessment, and forensic evaluation.The protocols are intended to assess generated images across visual realism, controllability, expert assessment, and forensic utility.Source: §I · p. 2
  • Generated images show high fidelity in band structure and visual style, can fool experts under blind inspection, and are not reliably distinguished by existing detectors and screening platforms.The authors present them as challenging controlled cases for validating and developing WB forensic methods.Source: §I · p. 2

II. RELATED WORK

Prior WB forensics largely addresses conventional image manipulations, while AI-generated WB images and controllable synthesis remain less systematically studied. GAN-Blot frames synthesis as separate control of structure and style, without relying on predefined appearance labels.

II. RELATED WORK

  • Existing WB forensic methods primarily target splicing, copy-move, duplication, and image reuse, while studies of AI-generated WB images are only beginning.The cited work frames its focus as controllable synthesis for constructing images to validate WB forensic methods.Source: §II · pp. 2–3
  • Prompt-guided and attribute-supervised synthesis is difficult to apply because WB structures lack standardized semantic entities and appearance annotations.Lanes and bands do not correspond to the semantic entities used by natural-image prompts, and WB lacks standardized labels such as pose or age.Source: §II · p. 3
  • Existing WB synthesis relies mainly on image-to-image translation without explicit control of lane configurations or band geometry, and lacks a benchmark for controllability and forensic utility.GAN-Blot instead formulates synthesis as structure–style conditional generation, providing stronger structural controllability than direct I2I translation.Source: §II · p. 3

A. Overview

GAN-Blot separates protein-band structure from gel or membrane appearance, using a mask and a style reference as distinct inputs. Its training recombines these conditions and uses a dual-path scheme to learn synthesis while anchoring it to real WB images.

A. Overview

  • A band-contour mask specifies local band locations, shapes, and lane layouts, while a style-reference image specifies global appearance such as background intensity, noise, texture, and lane-trace clarity.The formulation gives appearance explicit guidance rather than mapping a binary mask directly to one implicitly chosen grayscale appearance.Source: A. Overview · p. 3
  • Training randomly pairs a real WB image and its mask with a style reference, allowing N masks and M styles to form up to N × M combinations.The generator receives the mask and the style code extracted from the reference image.Source: A. Overview · p. 3
  • The dual-path scheme trains synthesis for structural consistency and style alignment, while self-supervision reconstructs real WB images from their extracted masks and style codes.The reconstruction path anchors recombined structure–style conditions to photorealistic WB appearance.Source: A. Overview · p. 3, Figure 3 · p. 4
  • The authors choose StyleGAN2 because its backbone directly injects reference-style information; they note diffusion training needs more domain-specific data and iterative sampling raises combination-generation cost.The paper states these considerations in the context of limited WB training samples.Source: A. Overview · pp. 3–4

B. Subnetworks

GAN-Blot combines mask-conditioned structure synthesis with reference-guided style encoding, then uses complementary losses to preserve band geometry, match appearance, and reconstruct real WB images.

B. Subnetworks

  • The StyleGAN2-based synthesizer conditions generation jointly on a protein-band mask and a style-reference image, while the segmentor encodes band shape without horizontal skip connections.The style encoder uses pretrained CLIP features as a reference-guided appearance descriptor, without predefined WB appearance labels or text.Source: B. Subnetworks · pp. 4–5

C. Loss Functions

  • A PatchGAN discriminator promotes local realism, and the method uses relativistic pairing GAN loss because overlapping patches violate WGAN-GP’s input-independence assumption and cause training instability.The section identifies binary cross-entropy as the band-shape loss measure.Source: B. Subnetworks · p. 4, C. Loss Functions · p. 5
  • Synthesis-path losses preserve band shape and align the output with the reference through histogram, CLIP style-code, and texture-consistency terms.Histogram alignment targets global brightness and contrast; texture consistency is computed from high-level features.Source: C. Loss Functions · p. 5
  • The self-supervision path uses reconstruction and band-shape losses to reconstruct each real WB input and maintain consistency between its input and segmented masks.The reconstruction loss encourages the segmented mask to serve as a 2D structural code for autoencoding the input.Source: C. Loss Functions · p. 5

D. Total Loss and Model Optimization

The model combines adversarial, structural, reconstruction, and style objectives, with empirically selected loss weights and alternating discriminator and generator-segmentor updates.

D. Total Loss and Model Optimization

  • The reported synthesizer loss expression includes band, histogram weighted 10, and code and style terms weighted 0.1 each; the segmentor expression includes band, reconstruction weighted 10, and 10LG.Source: D. Total Loss and Model Optimization · p. 5
  • Each iteration updates the discriminator first, then jointly updates the synthesizer and segmentor; R1 and R2 regularize discriminator gradients on real and generated data.The paper states these regularization terms are used to improve convergence stability.Source: D. Total Loss and Model Optimization · p. 5
  • Loss weights were tuned empirically by visual quality, one newly introduced term at a time, and training used one mask-image pair plus a random style reference per batch for approximately 1000K iterations.Source: D. Total Loss and Model Optimization · p. 5

IV. EXPERIMENT RESULTS

The experiments assess GAN-Blot’s controllability, reconstruction fidelity, and forensic performance, using a synthetic dataset and comparisons with generative models, detectors, platforms, and expert judgment.

IV. EXPERIMENT RESULTS

  • The study evaluates structural and appearance controllability, target-conditioned resynthesis, forensic detection, and expert visual assessment.Source: §IV · p. 5
  • WesternBlot46K comprises 45,861 GAN-Blot-generated images and an authentic subset of 151 cropped images plus 195 full-resolution laboratory WB photos.The 4,939 non-AI-generated WB images used for model development are not part of this dataset description.Source: §IV · p. 6
  • Comparisons include GAN-based and text- or instruction-guided generators, mask-conditioned image-to-image models, public and WB-specific detectors, screening platforms, and domain experts.The screening platforms named are Proofig and ImageTwin.Source: §IV · p. 6

A. Dataset Preparation

Dataset preparation combines publication-derived WB images for model development with a released synthetic-and-authentic research dataset and documented preprocessing and annotation procedures.

A. Dataset Preparation

  • The development corpus contains 4,939 WB images from 84 retracted or publicly questioned papers; 4,000 train the model and 939 are held for testing.The training set permits up to 16 million random combinations, and testing also includes 151 authentic cropped laboratory images.Source: A. Dataset Preparation · p. 6
  • The collected images span biomedical domains, journal tiers, and 2001–2019, while preprocessing removes labels and surrounding text and pairs images with thresholded, manually inspected band masks.Masks with evident contour errors are excluded, but minor noise or irregular boundaries are retained to reflect ambiguous weak-band boundaries.Source: A. Dataset Preparation · p. 6
  • The released WesternBlot46K dataset contains 45,861 synthetic images and authentic laboratory images, but not publication-cropped training images; access is for registered academic, non-commercial research.The paper contrasts it with public datasets containing only small re-synthesized patches.Source: A. Dataset Preparation · p. 6

B. Experimental Protocols and Compared Methods

The evaluation uses four protocols to test controllability, target-specific reconstruction, and forensic utility, alongside an expert visual assessment. GAN-Blot is compared with generative baselines and conditioned StyleGAN2 variants.

B. Experimental Protocols and Compared Methods

  • Four protocols assess structural control, appearance control, target-specific reconstruction fidelity, and whether forensic tools distinguish synthesized from non-synthetic WB images.P1 varies masks with a fixed style reference; P2 varies style references with a fixed mask; P3 evaluates resynthesis against a target; P4 tests forensic tools.Source: B. Experimental Protocols and Compared… · p. 6
  • Figure 5 arranges input protein-band masks in the leftmost column and style-reference images across the top row.The examples illustrate independent control of band structure and appearance style.Source: Figure 5 · p. 7
  • Figure 6 compares StyleGAN3, ADM, InstructPix2Pix, and ControlNet; noise-driven inputs lack structure or style control, while text- or instruction-based outputs are unreliable for WB synthesis.The caption attributes the latter limitation to missing well-defined semantic appearance descriptions and the lack of explicit style-reference inputs.Source: Figure 6 · p. 7
  • Forensic evaluation treats GAN-Blot outputs as synthetic and authentic laboratory or non-AI-generated publication-cropped WB images as non-synthetic.The tested tools include natural- or face-image detectors, WB-specific detectors, and publisher screening platforms.Source: B. Experimental Protocols and Compared… · pp. 6–7
  • The compared baselines include GAN-, diffusion-, text- or instruction-guided, and image-to-image translation models, with conditioned StyleGAN2 variants added for P3.StyleGAN2-J uses the band mask; StyleGAN2-JS uses both the mask and style reference, sharing GAN-Blot’s generator backbone.Source: B. Experimental Protocols and Compared… · p. 7
  • Expert visual assessment tests whether synthesized images appear plausible under human inspection.Experts inspect images without knowing the generation method; the assessment also considers whether structure and appearance are plausible.Source: B. Experimental Protocols and Compared… · p. 7

C. Controllability of WB Synthesis

GAN-Blot uses separate band-contour masks and style references to control WB structure and appearance. The results show that each input can guide its intended image properties while preserving the other.

C. Controllability of WB Synthesis

  • With a fixed style reference, GAN-Blot preserves the band locations, shapes, lane layouts, and empty-lane regions specified by different masks.A mask with no protein-band regions does not produce unintended bands, indicating that foreground structure is primarily guided by the input mask.Source: C. Controllability of WB Synthesis · pp. 7–8
  • With a fixed mask, GAN-Blot adapts background intensity, contrast, lane-trace visibility, noise, and gel or membrane texture to different style references while preserving band structure.The results show separate control of structure by the mask and appearance by the reference.Source: C. Controllability of WB Synthesis · p. 8
  • Noise-driven GAN and diffusion models lack explicit structural conditioning, while text- or instruction-guided models often fail to preserve band structure and realistic WB backgrounds together.The cited comparisons attribute instability to WB appearance lacking standardized semantic descriptions.Source: C. Controllability of WB Synthesis · p. 8
  • GAN-Blot combines a band mask for structural guidance with a WB style reference for appearance guidance under WB-specific visual constraints.The comparisons suggest that general-purpose generative models cannot reliably provide controllable WB synthesis directly.Source: C. Controllability of WB Synthesis · p. 8

D. Target-Conditioned Resynthesis and Loss Ablation

GAN-Blot’s target-conditioned resynthesis outperforms the compared methods on fidelity and band-structure measures, and experts rate its images highest across the reported plausibility measures. The loss ablations and assessments clarify the roles of training components and evaluator expertise.

D. Target-Conditioned Resynthesis and Loss Ablation

  • GAN-Blot achieves the best MSE, PSNR, SSIM, LPIPS, and IoU among the compared resynthesis methods, preserving target appearance and input band structure.CycleGAN lacks explicit mask conditioning, while StyleGAN2-JS uses the same conditioning inputs and generator backbone but has substantially lower resynthesis fidelity.Source: D. Target-Conditioned Resynthesis and… · p. 8
  • Figure 7 shows GAN-Blot reproducing target band locations, shapes, lane patterns, background intensity, and gel or membrane texture; baselines distort structure or mismatch appearance.The examples correspond to the P3 reconstruction results.Source: D. Target-Conditioned Resynthesis and… · pp. 8–9, Figure 7 · p. 9
  • Removing histogram loss degrades resynthesis fidelity, while removing band-shape losses reduces structural consistency with the input mask.Histogram loss supports global appearance, including background brightness and contrast.Source: D. Target-Conditioned Resynthesis and… · p. 9
  • Removing style-code or style-consistency losses weakens reference-guided appearance control, and removing reconstruction loss degrades resynthesis fidelity.The ablation results attribute these functions to the CLIP-based style-code constraint, background-texture constraint, and self-supervision path, respectively.Source: D. Target-Conditioned Resynthesis and… · p. 9

E. Expert-based Visual Assessment

  • GAN-Blot receives the highest expert-rated fooling rate, structure plausibility, style plausibility, and realism rate among the compared methods.StyleGAN2-JS performs better than CycleGAN, ControlNet, and InstructPix2Pix but remains below GAN-Blot.Source: E. Expert-based Visual Assessment · pp. 9–10, Table II · p. 9
  • Manjunath et al.’s detector separates its official real and fake subsets but does not reliably respond to GAN-Blot images and has unstable responses to authentic laboratory WB images.Figure 8 presents score distributions for the official subsets, GAN-Blot samples, and cropped authentic laboratory images.Source: Figure 8 · p. 10
  • Image-processing engineers’ fooling-rate judgments differ from domain experts’, with engineers often overlooking implausible WB structures or appearance in CycleGAN and ControlNet outputs.For StyleGAN2-JS, engineers report a lower fooling rate than domain experts.Source: E. Expert-based Visual Assessment · p. 10

F. Forensic Evaluation under Existing Detectors

The evaluation finds that existing AI-image detectors, WB-specific synthetic-image detectors, and publisher-used screening platforms do not reliably identify GAN-Blot images. WesternBlot46K is intended to support validation and development of WB forensic methods, while dedicated detection of experiment-inconsistent WB evidence remains open.

F. Forensic Evaluation under Existing Detectors

  • General-detector evaluation used two GAN-Blot subsets and three non-AI WB subsets to assess synthetic-image detection and false alarms.The synthetic subsets comprised manually inspected images and randomly selected images; the non-AI subsets covered cropped and full laboratory images and publication-cropped images.Source: F. Forensic Evaluation under Existing… · p. 10
  • All tested general AI-image detectors had below 20% TPR on manually inspected GAN-Blot images, and their WB-domain false-alarm behavior was unstable.On randomly selected GAN-Blot images, only RINE reached about 30% TPR; Effort and UnivFD also produced non-negligible false positives on non-AI WB images.Source: F. Forensic Evaluation under Existing… · p. 10
  • 49.51% median BACC across 120 Cardenuto configurations; Manjunath’s detector also did not reliably respond to GAN-Blot or authentic laboratory images.Only nine Cardenuto configurations exceeded 55% BACC, and only four exceeded 60%.Source: Figure 9 · p. 11, F. Forensic Evaluation under Existing… · p. 11
  • Both ImageTwin and Proofig reported 0 AI-generated images for the representative GAN-Blot images submitted, in a limited supplementary screening.The authors describe this as representative-example screening, not a large-scale platform benchmark.Source: F. Forensic Evaluation under Existing… · p. 11
  • WesternBlot46K is intended to support WB-forensics research through controllable images with specified band structures and background appearances.Developing dedicated detection of experiment-inconsistent WB evidence remains open and requires additional expert annotation of generated images.Source: F. Forensic Evaluation under Existing… · p. 11

V. CONCLUDING REMARKS

GAN-Blot formulates WB synthesis as structure–style conditional generation and establishes a benchmark for WB forensic research. The generated images are realistic yet challenging for existing forensic tools, providing controlled cases for validating and developing WB forensic methods.

V. CONCLUDING REMARKS

  • GAN-Blot independently controls band geometry and appearance characteristics without predefined attribute annotations.The benchmark includes a large-scale synthetic WB dataset and protocols for controllability, target-conditioned resynthesis, expert visual assessment, and forensic evaluation under existing detectors.Source: §V · p. 11

Related papers

Loading 2609.06619v1…