Geng Academic Fraud Detection Report: "Unsupervised Visual Anomaly Detection Using Self-Supervised Pre-Trained Transformer" (DOI: 10.1109/ACCESS.2024.3454753)

Academic fraud report · Geng Detector

Summary

Verdict: Cleared (✅). This IEEE Access 2024 paper by Jun-Hyung Kim and Goo-Rak Kwon was reviewed under the Geng six-style framework focused on data logic, timeline consistency, and experimental plausibility, since standard biomedical image-fraud checks do not apply. Key checks all passed: (1) Timeline and citation consistency are clean — DINOv2 references [25] (arXiv:2304.07193) and [26] (arXiv:2309.16588) are correctly dated 2023, and the paper's 14 August 2024 submission date respects that chronology; the 14×14 patch specification matches DINOv2's official configuration. (2) Hardware/time plausibility is reasonable: an RTX 3090 running ViT-g for 300 epochs producing ~354.9 minutes of training time on the Bottle category is physically credible. (3) Ablation tables show logically monotonic trends across model size (ViT-B → ViT-L → ViT-g), masking ratio (10% best at 86.38, 40% drops to 85.72), and incremental synthetic anomaly types (79.13 → 82.84 → 86.38). No suspiciously perfect deviations. The only flagged observation is the journal's 19-day submission-to-acceptance window, which is consistent with IEEE Access norms rather than author misconduct. Pixel-level image analysis was not possible due to lack of raw figures. Confidence: high for textual analysis; image-level forgery not assessable.

Verdict

Cleared (✅). No substantive evidence of academic fraud was identified. The paper passes timeline, citation, hardware-cost, and ablation-logic checks. The sole flag — a 19-day acceptance window — reflects IEEE Access's standard editorial pace, not author misconduct.

Key findings

Timeline and citation consistency: Clean. DINOv2 citations correspond to 2023 arXiv preprints (2304.07193 and 2309.16588), and the paper's submission on 14 August 2024 postdates them appropriately. The 14×14 patch parameter matches DINOv2's official specification.
Publication speed flag: Received 14 August 2024, accepted 2 September 2024 — a 19-day turnaround. This is characteristic of IEEE Access's rapid-review model and not indicative of author-side issues, though it may limit depth of peer review.
Hardware and compute plausibility: An RTX 3090 (released 2020) is a reasonable hardware choice for 2024 experiments. Reported training time of ~354.9 minutes on the Bottle category (6 blocks, 10 heads) for ViT-g with 300 epochs is physically credible.
Ablation logic: Monotonic trends observed in Table 5 (model size scaling), Table 6 (masking ratio with 10% optimal at 86.38, declining to 85.72 at 40%), and Table 8 (incremental synthetic anomaly types: 79.13 → 82.84 → 86.38). No unrealistically perfect statistics.
Image-level analysis: Not performed — raw high-resolution figures were unavailable for ELA, noise, or duplicate-image checks. No visible anomalies in qualitative figures (e.g., Figure 4 heatmaps) based on text description.

Evidence highlights

DOI: 10.1109/ACCESS.2024.3454753
Submission: 14 August 2024; Acceptance: 2 September 2024 (19 days)
Reference [25]: arXiv:2304.07193 (April 2023); Reference [26]: arXiv:2309.16588 (September 2023)
DINOv2 patch configuration reported as 14×14 — consistent with official model
Training hardware: NVIDIA GeForce RTX 3090
Table 6 masking ratio: 10% → 86.38; 40% → 85.72
Table 8 synthetic anomaly ablation: 79.13 → 82.84 → 86.38
Table 5 model scaling: ViT-B → ViT-L → ViT-g trend

Notes

Limitations: Pixel-level image forgery detection (ELA, EXIF, layer analysis) was not feasible due to absence of original figure files. Findings are confined to textual logic, citation integrity, and experimental reasonableness.
The 19-day review window at IEEE Access, while typical for the venue, warrants reader caution regarding peer-review depth for any paper in this outlet.
No contact with authors or journal editorial action is recommended at this time.
This automated assessment does not replace formal institutional investigation; false positives and false negatives are possible.