Academic Fraud Detection Report: 'Zero-Shot Industrial Anomaly Detection via CLIP-DINOv2 Multimodal Fusion and Stabilized Attention Pooling' (Electronics, 2025)

Academic fraud report · Geng Detector

Summary

This report examines a paper published in Electronics (MDPI), DOI 10.3390/electronics14244785, dated December 2025. The verdict is confirmed (实锤) due to serious logical and chronological anomalies. The most critical finding is that the paper, received October 28 2025, cites four journal articles purportedly published in 2026 (Expert Systems with Applications volumes 296–297 and Advanced Engineering Informatics volume 69 Pt B). The detection of multiple citations to future-dated sources constitutes a clear logical contradiction. Secondary findings include an unusually fast editorial cycle (revision to acceptance in 2 days; submission to publication in 38 days), and a high inference-time standard deviation (±18.3 ms on a 43.4 ms mean) suggesting poor benchmarking rigor. The author also reported that mean values in Tables 1–2 were arithmetically consistent. Limitations: original high-resolution figures were not available for image-duplication analysis; conclusions on figures remain pending. Confidence in the time-travel citation issue is high; confidence in the editorial-speed claim is lower due to lack of insight into MDPI internal processes.

Verdict

Confirmed (实锤): The paper contains a clear temporal paradox — four references are dated 2026 while the manuscript was submitted in October 2025. Additional concerns exist regarding editorial speed and benchmark methodology, but the citation issue alone is sufficient to raise serious questions about scholarly rigor.

Key findings

Time-travel citations: At least four references are dated 2026 in a manuscript received 28 October 2025 and published 5 December 2025.
Ref 1: *Expert Syst. Appl.* 2026, 297, 129489
Ref 3: *Expert Syst. Appl.* 2026, 296, 128965
Ref 6: *Expert Syst. Appl.* 2026, 296, 128995
Ref 7: *Adv. Eng. Inform.* 2026, 69 Pt B, 103931
Rushed editorial handling: Received 28 Oct 2025 → Revised 24 Nov 2025 → Accepted 26 Nov 2025 → Published 5 Dec 2025. Revision-to-acceptance interval of 2 days; total submission-to-publication of 38 days. This pattern is consistent with special-issue fast-track or outsourced production, though this reflects editorial practice rather than direct author misconduct.
Benchmarking irregularity: Inference time reported as 43.4 ± 18.3 ms per image on RTX 4090. A relative standard deviation of ~42% is unusually high for a warm-started GPU inference benchmark and suggests either missing warm-up procedures or a very small sample size.
Table arithmetic verified (no anomaly detected): Re-computation of reported means matched the published values.
Table 1 (Image-AUROC mean): (94.8 + 84.6 + 84.1 + 96.4 + 98.0 + 99.0 + 97.0) / 7 = 93.414% → reported 93.4% ✓
Table 2 (Image-AP mean): (97.8 + 87.6 + 86.4 + 98.0 + 94.6 + 96.9 + 98.9) / 7 = 94.314% → reported 94.3% ✓

Evidence highlights

Citation chronology is the strongest direct evidence and is internally verifiable from the paper's reference list and submission metadata.
Editorial timestamps are taken from the paper's first-page history (Received / Revised / Accepted / Published dates) and represent standard MDPI metadata.
Inference-time statistics are taken from Section 4.1.3 (Implementation Details).
Mean-value re-checks are reproducible from Tables 1 and 2.

Notes

Figures 1–12 could not be analyzed for image reuse or Photoshop artifacts because high-resolution source images were not provided. A pixel-level forensics pass is recommended when figures become available (e.g., via PubPeer or the journal's supplementary files).
The time-travel citation issue may in principle be explained by early online / preprint versions or accepted-but-not-yet-typed references; however, the simultaneous appearance of four such citations materially reduces the plausibility of an innocent explanation and warrants author clarification.
All numeric evidence (volumes, article numbers, percentages, dates, inference timings) is preserved exactly as reported. No additional findings have been invented.
Limitations: this report relies on the textual content provided; institutional, authorship-network, and image-based analyses remain pending.