Academic Integrity Review: HE-DeepFM: An FHE Inference System for CTR Prediction with Efficient FM Interactions (SIGIR '26)

Academic fraud report · Geng Detector

Summary

This review assesses potential research misconduct in the SIGIR '26 paper 'HE-DeepFM' (DOI: 10.1145/3805712.3809866). Verdict: SUSPICIOUS-INCONCLUSIVE. Four review areas were examined: (1) Mathematical derivation and algorithmic logic in Section 3.2 and Algorithm 1 were verified as self-consistent; the O(F²) feature interaction reduction via the identity (Σu)² − Σu² is mathematically sound and implemented via AcrossBlockSum/WithinBlockSum in CKKS SIMD slots. (2) Experimental data in Table 2 (latency breakdown) was internally consistent: e.g., for HE-DeepFM HE-FM at Threshold=500, Emb(2.5)+Bot(1.4)+Top(7.9)+Boot(49.0)=60.8s vs. reported 60.7s (rounding), and speedup 172.9/60.7≈2.85× matches. The bootstrapping-dominant profile (70–90% runtime) is realistic for FHE systems. (3) Citation timeline is coherent with a 2026 venue: Orion (Ebel et al., 2025), HE-LRM baseline (Garimella et al., 2025), and Intel Xeon Platinum 8488C + NVIDIA A100 hardware are all time-appropriate. (4) No pixel-level image analysis was possible because only text/tabular data was available. No fabrication indicators were detected; limitations are explicitly noted.

Verdict

🟡 Suspicious – Inconclusive (no evidence of misconduct found; insufficient data to fully clear)

The paper passed all text- and data-based integrity checks. The sole unresolved limitation is the inability to perform pixel-level analysis of figures due to lack of high-resolution image assets.

Key findings

Mathematical derivation verified. The identity (Σu)² − Σu² correctly reduces O(F²) pairwise interactions and is rigorously implemented via AcrossBlockSum / WithinBlockSum in CKKS SIMD slots (Section 3.2, Algorithm 1, Figure 2).
Table 2 latency arithmetic is internally consistent. HE-DeepFM HE-FM at Threshold=500: Emb (2.5s) + Bot (1.4s) + Top (7.9s) + Boot (49.0s) = 60.8s vs. reported Total 60.7s; speedup 172.9 / 60.7 = 2.85× matches the paper's claim.
Performance profile is realistic. Bootstrapping dominates 70–90% of runtime, consistent with known FHE bottlenecks—not a textbook-perfect artificial signature.
Citation timeline is coherent for a 2026 venue. Orion (Ebel et al., 2025), HE-LRM baseline (Garimella et al., 2025), and Intel Xeon Platinum 8488C + NVIDIA A100 hardware are all temporally appropriate.
No image-level analysis performed. Figures 1–5 could not be inspected for visual duplication, PS artifacts, or noise-pattern anomalies.

Evidence highlights

| Check | Numeric evidence | Outcome | |---|---|---| | Sum check (Table 2, Threshold=500) | 2.5 + 1.4 + 7.9 + 49.0 = 60.8 vs. reported 60.7 | ✅ Consistent within rounding | | Speedup check | 172.9 / 60.7 = 2.849 ≈ 2.85× | ✅ Matches paper | | Reference recency | Orion 2025, HE-LRM 2025 | ✅ Timeline plausible | | Hardware | Intel Xeon Platinum 8488C (Sapphire Rapids) + NVIDIA A100 | ✅ Appropriate for 2026 | | Image forensics | Not performed | ⚠️ Inconclusive |

Notes

DOI: 10.1145/3805712.3809866
Conference: SIGIR '26 (20–24 July 2026); review conducted 2026-06-22.
Limitations: pixel-level figure forensics, dataset-level replication, and code-availability verification were not possible from the supplied materials.
No action recommended against the authors based on available evidence; optional visual verification of figures by readers with access to the original PDF is advisable for 100% clearance.
This report is AI-assisted and intended for academic discussion only; final misconduct determinations require institutional investigation.