Geng Academic Integrity Report: Feature Enhancement Module Based on Class-Centric Loss for Fine-Grained Visual Classification (IEEE TNNLS, DOI 10.1109/TNNLS.2025.3613791)

Academic fraud report · Geng Detector

Summary

This report assesses the paper 'Feature Enhancement Module Based on Class-Centric Loss for Fine-Grained Visual Classification' (IEEE TNNLS, DOI: 10.1109/TNNLS.2025.3613791) by Wang et al. (East China Normal University). Overall verdict: highly suspicious (🟠/🔴), though with important caveats. Key findings: (1) t-SNE visualizations (Figures 3 and 8) show implausibly perfect, noise-free cluster separation, consistent with cherry-picking or manipulation, though benign explanations cannot be excluded; (2) a confirmed internal contradiction in the publication metadata: the 'date of current version' (4 February 2025) precedes the received date (7 February 2025) — most likely a typesetting error; (3) automated image forensics detected copy-move matching blocks and noise inconsistencies across many figures, but multi-panel layout structure may produce false positives; (4) numerous broken cross-references, truncated sentences, and an anomalous citation (Darwin 1895), suggesting PDF extraction or referencing tool issues; (5) all performance tables (Tables I–X) were missing from the provided text, preventing numeric verification. The automated Bayesian synthesis (≥99.9% posterior) should be treated cautiously given weak individual evidence. Formal misconduct determination requires institutional investigation and original data from the authors.

Verdict

Highly suspicious (🟠 in main report; automated synthesis rates 🔴 with posterior ≥99.9%) — but this assessment is subject to significant limitations. The paper is a deep learning/computer vision work; no biological imagery (Western blots, gels) is involved, and all data tables (Tables I–X) were missing from the provided text, making GRIM/numeric forensic checks impossible. The Bayesian posterior (BF = 5.6e+38) assumes conditional independence of evidence and should be interpreted with caution; no single finding is individually conclusive.

Key findings

⚠️ Implausibly perfect t-SNE clustering (Figures 3, 8): method-side clusters show near-perfect separation with no overlap, outliers, or boundary blur across 20 CUB classes; CE-loss baseline shows visible structure. t-SNE typically yields overlap and outliers even at high accuracy. Possible benign explanation: cherry-picking of random seeds (selective presentation rather than fabrication); no seed or projection-selection procedure is documented.
✅ Confirmed publication timeline contradiction (p. 893 footnote): "date of current version 4 February 2025" precedes the received date (7 February 2025), revised date (30 June 2025), and acceptance date (21 September 2025). Verified verbatim. Most likely a typesetting error (should be 2026), not evidence of fraud, but an anomaly requiring correction.
⚠️ Image forensics flags: copy-move detection across many figures (e.g., img-019.png offset (32,0), 230 matched block pairs; img-015.png offset (0,32), 134 pairs; img-000.jpg/img-013.png/img-025.jpg noise variance CV≈1.2–1.4). Multi-panel layouts (5×3 Grad-CAM grids, dual-panel t-SNE) can produce same-offset false positives, and "Authorized licensed use" watermarks/secondary compression can explain noise unevenness; however, no affirmative statement in the paper supports these benign explanations.
⚠️ Text integrity anomalies: missing figure numbers ("Fig.."), dangling citations (",,"), truncated sentences, and reference [41] (Darwin 1895, *Proc. Roy. Soc. London*, vol. 58, pp. 240–242) — verified present and contextually incongruous. Likely PDF extraction/typesetting issues or citation-manager artifacts; not fraud evidence per se.
ℹ️ Tables I–X absent from provided text: numeric self-consistency unverifiable (e.g., the claim that Pearson similarity outperforms cosine on CUB, CAR, AIR, LOC while cosine wins marginally on NAB cannot be checked). Scattered in-text values ("90.72% to 91.52%", "+2.36% on NAB", "95.44% with CE") were verified to exist in the text. This is a capability limitation, not an accusation.

Evidence highlights

Verified verbatim footnote contradiction: current version date (2025-02-04) earlier than received (2025-02-07), revised (2025-06-30), accepted (2025-09-21) dates.
Copy-move detections: img-035.png offset (0,32), 58 pairs; img-021.png offset (32,0), 96 pairs; img-010.jpg offset (40,0), 49 pairs; img-023.png offset (32,0), 60 pairs; img-005.jpg offset (0,40), 38 pairs (each logLR=1.46).
Visual analysis independently noted Figures 3 and 8 "lack realistic data noise" (medium confidence).
Figure 7: inconsistency between background texture of the original-image row and heatmap row (visual summary), potentially related to heatmap base-image substitution — unconfirmed.
Automated Bayesian synthesis: posterior ≥99.9%, 95% CI [100%, 100%], BF = 5.6e+38; prior fixed at 5%, conditional independence assumed — likely overstates joint strength of correlated evidence.

Notes

Recommended follow-up: request t-SNE raw feature files, code, and random seeds; request Figure 7 base images and Grad-CAM source files; ask authors/journal to correct the footnote date; post technical questions on PubPeer; submit forensic materials to IEEE TNNLS editors; check GitHub repo (Richard5413/FEM-CC) against the paper.
The benign explanations for image-forensic findings are generic speculation, not supported by affirmative statements in the paper; the actual locations of matched blocks cannot be determined from available information.
AI-assisted report for academic discussion only; formal misconduct determination requires professional institutional investigation. False positives and false negatives are possible.