Integrity Review: "UniAD: A Real-World Multi-Category Industrial Anomaly Detection Dataset with a Unified CLIP-Based Framework" (Information, MDPI, 2025)

Academic fraud report · Geng Detector

Summary

This review reports on the paper UniAD (DOI: 10.3390/info16110956), published in Information (MDPI) in November 2025. The verdict is highly suspicious, primarily due to the presence of unrevised reviewer-comment text within table captions. Specifically, the captions of Table 10 (Page 19) and Table 14 (Page 21) reproduce verbatim what appear to be reviewer critiques rather than academic descriptions, suggesting an unrevised draft was submitted. A second major concern is the Data Availability Statement, which confirms the UniAD dataset (25,000+ images) is non-public due to industrial confidentiality, undermining the paper's value as a benchmark. A potential statistical irregularity is also flagged: in Table 14, increasing epochs from 5 to 10 at LoRA Rank 8 changes AUROC only marginally (92.1 → 92.0), a pattern suggestive of hyperparameter tuning. Finally, Table 8 shows DSR achieving only 5.4% F1 yet 75.0% AUROC on RESISTOR, indicating likely baseline implementation issues. These findings are interpretive; confirmation requires editorial investigation.

Verdict

Highly Suspicious. The manuscript contains unresolved reviewer-comment text in Table 10 and Table 14 captions, and the central dataset (UniAD, 25,000+ images) is not publicly released, severely limiting reproducibility and the validity of its benchmark claims.

Key findings

Unrevised reviewer text in table captions: Table 10 (Page 19) and Table 14 (Page 21) reproduce passages that read as referee critiques rather than academic descriptions, indicating the authors uploaded a non-final draft.
Table 10 caption: *"Optimality and efficiency issues need to be questioned about the adaptation mechanism itself. While LoRA provides parameter efficiency, the authors never speak to why exactly this particular adaptation mechanism is optimal for anomaly detection."*
Table 14 caption: *"Illustrative examples of anomalies and CLIP embeddings are sparse; add visualizations of attention maps or anomaly heatmaps on UniAD samples to elucidate framework mechanics."*
Non-public benchmark dataset: The Data Availability Statement (Page 23) explicitly states: *"The data are not publicly available due to confidentiality agreements with the industrial partners."* This prevents independent verification of the reported SOTA results.
Suspicious ablation pattern (Table 14): At LoRA Rank 8, AUROC is 92.1 at 5 epochs and 92.0 at 10 epochs; F1 shifts from 89.1 to 89.2. Epoch scaling shows near-flat performance rather than the expected convergence/overfit signature.
Implausible baseline result (Table 8 vs. Table 9): DSR on RESISTOR yields AUROC = 75.0% but F1 = 5.4%, an unusually large gap suggesting a baseline implementation or thresholding bug, which may unfairly advantage the proposed method (reported F1 = 99.1%).

Evidence highlights

DOI: 10.3390/info16110956
Table 10 (Page 19): Captions reproduced verbatim from reviewer-style critique.
Table 14 (Page 21): Same pattern; caption reads as referee comment.
Table 14 ablation numbers: LoRA Rank 4/8/16 at 5 epochs → AUROC 91.8 / 92.1 / 91.9; Rank 8 at 10 epochs → AUROC 92.0.
Table 8 / Table 9: DSR on RESISTOR — F1 = 5.4%, AUROC = 75.0%; the proposed method claims 99.1% F1.
Data Availability Statement (Page 23): Explicit non-public declaration citing industrial confidentiality.

Notes

Findings are based on textual and tabular inspection; no image or code analysis was performed.
The 0.1-point AUROC variation in Table 14 is small and within typical noise; flagged as a pattern concern rather than definitive evidence.
The DSR anomaly (75.0% AUROC vs. 5.4% F1) is consistent with a thresholding or implementation bug rather than data fabrication, but it does raise reproducibility concerns.
The unrevised reviewer text in table titles is the strongest indicator of poor editorial control; this should be verified against the publisher's version of record.
Final adjudication of misconduct requires investigation by the journal editor or institutional ethics body.