Academic Fraud Detection Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

Verdict: Substantiated (实锤). The reviewer's analysis identified multiple high-confidence data-integrity issues in a 2025 ACM MM paper (DOI: 10.1145/3746027.3755336). The primary finding concerns Table 1 (linear classification on CheXpert): the proposed method's AUC values at 1% and 10% training data are identical (89.5, 89.5), which is statistically implausible given a 10× increase in training data. Furthermore, the reported gains over the MRM baseline form a constant +1.0 offset across all data regimes, strongly suggestive of manually adjusted figures rather than genuine experimental results. A second finding flags a self-contradictory ablation study (Table 4): adding the PAR module to DKBA decreases CheX-ray14 AUC from 81.0 to 80.8, yet the text in Section 5.3.1 claims complementary benefits. Image reuse checks were not performed due to lack of high-resolution source files. Confidence is high for findings 1–2; figure-level verification remains limited. The report does not constitute a formal finding of misconduct.

Verdict

Substantiated concerns (🔴 实锤). The report identifies two reproducible, textually evident anomalies that together constitute strong prima facie evidence of data fabrication or manipulation. Figure-level verification was not possible within the available inputs.

Key findings

Suspiciously identical AUC values across data scales (Table 1, CXP / ViT-based). The proposed model reports AUC = 89.5 at both 1% and 10% training data. A 10× increase in training data producing no change at 0.1% precision is statistically implausible for stochastic deep-learning training.
Constant +1.0 offset versus MRM baseline (Table 1). MRM: 88.5 / 88.5 / 88.7; Ours: 89.5 / 89.5 / 89.7. The differences across the 1%, 10%, and 100% regimes are exactly +1.0, +1.0, +1.0, which is inconsistent with organic experimental variation.
Ablation contradicts textual claim (Table 4 vs. Section 5.3.1). Without PAR (DKBA only), CheX-ray14 AUC = 81.0. With PAR + DKBA (full model), AUC = 80.8. Adding PAR reduces performance by 0.2 AUC, yet the paper asserts the modules are "complementary" and presents the full model as SOTA.
Image-level reuse check: not conducted due to absence of high-resolution source images; pixel forensics (background noise, splicing artifacts) are therefore unavailable.

Evidence highlights

Table 1 — Linear classification results, CheXpert (CXP), ViT-based rows:
MRM: 88.5 / 88.5 / 88.7
Ours: 89.5 / 89.5 / 89.7
ΔOurs − MRM across regimes: +1.0, +1.0, +1.0 (exact match).
Table 4 — Ablation on CheX-ray14 (AUC):
DKBA only: 81.0
DKBA + PAR (full): 80.8 (decrease of 0.2)
Section 5.3.1 — Claims "complementary benefits" of PAR and DKBA, directly contradicted by the ablation row above.
DOI: 10.1145/3746027.3755336 (ACM MM '25). Venue dates 27–31 Oct 2025 are temporally consistent with a 2026 review date.

Notes

Confidence is high for findings 1–2 because they are derived from explicit numeric tables in the manuscript and require no image processing.
A definitive misconduct determination requires institutional investigation, including raw training logs, multi-seed variance, and code review.
Recommended actions from the original report: request raw logs and per-seed variance from the authors; raise a PubPeer comment focused on Table 1 and Table 4; notify the ACM MM program chairs; report to the corresponding author's institution (Chongqing University of Posts and Telecommunications).
This automated report is advisory only; false positives are possible, and the authors retain the right to respond with original experimental artifacts.