Geng Report: Adaptive Policy Learning for Connected Autonomous Vehicles Defending Malicious Access Requests by Graph Reinforcement Learning (DOI: 10.1109/JIOT.2024.3429522)

Academic fraud report · Geng Detector

Summary

This Geng academic-integrity report flags the paper 'Adaptive Policy Learning for Connected Autonomous Vehicles Defending Malicious Access Requests by Graph Reinforcement Learning' by Qian Xu, Lei Zhang, and Yixiao Liu (IEEE Internet of Things Journal, Vol. 11, No. 20, 2024; DOI: 10.1109/JIOT.2024.3429522) as highly suspicious. The central concern is Table V, where for the Interfering Access Requests task both GCN and STGCN yield identical Precision, Recall, and F1-Score values to four decimal places (GCN: 0.9691; STGCN: 0.9994), a near-impossible coincidence in binary classification that strongly suggests fabricated or post-hoc tuned numbers. Secondary issues include copy-paste artifacts in the GNN/GRL review (the same citation [13] is assigned to two unrelated tasks, and a dangling 'both used' clause), suspiciously fraction-like TRR values (98.48% and 93.94%, consistent with n=66) presented with two-decimal precision on a seemingly tiny test set, and a potential publication-date conflict in reference [1] cited from a March 2024 submission. Confidence is high for the textual anomalies and the Table V issue, while image-level checks (Figs. 7, 9, 14) were not feasible in this text-only review.

Verdict

🟠 Highly suspicious. The Table V coincidence is statistically implausible and indicates likely data fabrication or aggressive post-hoc tuning. Additional textual and methodological anomalies reinforce concern. Image forensics were not performed.

Key findings

Statistical impossibility in Table V (Interfering Access Requests): GCN Precision = Recall = F1 = 0.9691; STGCN Precision = Recall = F1 = 0.9994. In binary classification, identical Precision and Recall at four decimal places require FP = FN; this occurring for both compared models on the same test set is essentially impossible by chance.
Copy-paste / logic errors in Section II.C.2 (p. 33479): A single citation [13] is attributed first to 'collaborative lane-changing' and immediately after to 'signaling control'. The phrase 'Liu et al. [34] both used' appears with a stranded 'both', indicating unrevised reused text.
Suspicious sample size / TRR values (Section V.F.2, p. 33489): Reported TRRs of 98.48% and 93.94% match 65/66 and 62/66, suggesting a total of ~66 test samples; two-decimal precision on such an n is unjustified yet paired with a smooth Fig. 14 convergence curve.
Potential reference timeline issue: A 2024 Global Automotive Cybersecurity Report cited in a manuscript reportedly submitted on 12 March 2024 raises a plausible availability/date conflict.
Image checks not performed: Figures 7, 9, and 14 could not be analyzed at pixel level; absence of evidence here is not evidence of absence.

Evidence highlights

Table V (GCN vs STGCN comparison, Interfering Access Requests): 0.9691 / 0.9691 / 0.9691 and 0.9994 / 0.9994 / 0.9994.
Section II.C.2 quotation: 'Chen et al. [13] proposed a method based on GCN with DQN... for collaborative lane-changing... Chen et al. [13] proposed GCN and DQNs for signaling control.'
Section II.C.2 dangling 'both': 'Liu et al. [34] both used GCN as the state space...'
TRR values 98.48% (≈65/66) and 93.94% (≈62/66) consistent with n=66.
Submission date reported as 12 March 2024 vs reference [1] cited as a 2024 report.

Notes

DOI: 10.1109/JIOT.2024.3429522
Authors: Qian Xu, Lei Zhang, Yixiao Liu (Tongji University per report).
All numerical evidence above is reproduced exactly as reported; no independent reanalysis was performed.
Recommended actions from the report: request raw confusion matrices and test-set size for Table V, raise the issue on PubPeer; escalation to journal editor and institutional committee is pending.
This translation preserves the original report's claims and uncertainty markers; final determination of misconduct requires an official investigation.