Integrity Review: Joint Three-Task Optical Performance Monitoring with High Performance and Generalizability Using a Meta-Learning Algorithm in WDM Transmissions (ICICC 2024)

Academic fraud report · Geng Detector

Summary

This AI-assisted review examines the 2024 ICICC paper 'Joint Three-Task Optical Performance Monitoring...' (DOI: 10.1109/ICICC63565.2024.10780686) by Di Zhang et al. The overall verdict is 'highly suspicious' per the report's grading rules, driven by three independent minor anomalies and an automated Bayesian synthesis (posterior 91.5%, BF = 205.37). However, every individual finding carries a plausible benign explanation, and the report itself states there is no concrete evidence of misconduct. Key issues: (1) MFI accuracy is a perfect 100% across all modulation formats in both single-channel and WDM fine-tuning scenarios, which the authors attribute to distinct format features in purely simulated data; (2) error curves for OSNR and NLNP are extremely smooth with no outliers, plausibly explained by noise-free simulation data; (3) copy-move forensics flagged regular repeated structures in five figures, but detected offsets (multiples of 32/64 px) match multi-panel chart axes and grids rather than duplicated content; (4) a minor methodological ambiguity about train/test split vs. MAML support/query division. Confidence is limited: no original data, code, or high-resolution figures were available. The report recommends requesting raw data from the authors but considers journal or institutional escalation unnecessary at this stage.

Verdict

Highly suspicious (🟠), pending original data — per the report's re-review rule (3 retained independent anomalies → elevated rating). Critically, all three retained anomalies are minor and each has a stated benign explanation; the report explicitly finds *no concrete evidence* of misconduct. The automated Bayesian synthesis yielded a posterior of 91.5% (95% CI [59.9%, 98.5%], BF = 205.37) but flagged the highest single contributor as AI-generated text features (logLR = 1.97, 41.4% contribution), and 16 candidate leads were excluded after re-review as having benign explanations.

Key findings

Finding 1 (🟡, benign explanation noted): MFI task achieves a perfect 100% accuracy for all four modulation formats (QPSK/16QAM/32QAM/64QAM) in both the single-channel scenario (Fig. 4(a)) and the five-channel WDM fine-tuning scenario (Fig. 5(a)), despite ASE noise, 100 kHz laser linewidth, nonlinearities, and inter-channel interference. Authors state this is "due to the distinct feature differences between the various modulation formats" — plausible for purely simulated data, but zero errors across the full OSNR range and on only 20 samples/condition warrants verification with raw data.
Finding 2 (🟡, benign explanation noted): OSNR and NLNP error curves are extremely smooth and monotonic with no outliers across all four curves (Fig. 4(b)(c), Fig. 5(b)(c)). Reported improvements — OSNR mean error −18.9%, NLNP −18.4% (single-channel); OSNR −23.7%, NLNP −24.5% (WDM) — all verified against the text. Smoothness is plausible for noise-free simulation with 40 samples/group, possibly averaged.
Finding 3 (🟡, benign explanation noted): Copy-move forensics flagged 5 figures (img-002 to img-006), e.g., img-003: 108 pairs at offset (32,0); img-005: 95 pairs at (0,32); img-006: 72 pairs at (64,0). These are synthetic multi-panel bar/line charts (Fig. 4, Fig. 5, panels (a)(b)(c)) where shared axes, borders, and evenly spaced elements naturally produce same-offset matches at regular 32/64/96-pixel offsets — consistent with figure captions. Not photo-class evidence; no conclusion of data tampering. Requires high-resolution originals to exclude copied data points between panels.
Finding 4 (informational): Dataset arithmetic is self-consistent (single-channel: 40×4×16×9 = 23040; WDM: 20×4×16×6 = 7680), but the text's "evenly divided into training and testing sets" (11520 each) is not reconciled with the MAML support (5)/query (15) split — a writing ambiguity, not evidence of fraud; authors should clarify strict train/test isolation.

Evidence highlights

Automated synthesis: prior 5%, posterior 91.5%, 95% CI [59.9%, 98.5%], BF = 205.37; LR version geng_lr_v3 (n=11 calibration, shrinkage k=10).
Top contributors: [T1] AI-generated text features (logLR = 1.97, 41.4%); [Finding 3] copy-move detections (logLR = 1.19, 21.3%); [Finding 1] perfect MFI accuracy (logLR = 1.08, 19.4%); [Finding 2] overly smooth curves (logLR = 1.08, 17.9%).
16 candidate leads were excluded after review (benign explanations valid / test preconditions unmet / descriptive entries).
3 cited data points programmatically verified against source text; both dataset-size calculations verified.

Notes

All three retained anomalies are explicitly marked as having plausible benign explanations; no conclusive evidence of misconduct was identified.
Evidence was treated as conditionally independent in the synthesis, which may overstate joint strength of correlated evidence; the synthesis is auto-generated and not a finding of academic misconduct.
Recommended actions: request raw per-sample MFI predictions, train/test split scripts, and raw scatter data for error curves; verify high-resolution originals of Fig. 4/Fig. 5 to exclude panel-to-panel data-point copying (low priority). Optional PubPeer posting; journal/institutional escalation judged unnecessary at this stage.
Verified: DOI 10.1109/ICICC63565.2024.10780686; venue: 2024 4th International Conference on Intelligent Communications and Computing (ICICC).
Report is AI-assisted; final determination requires official investigation.