Integrity Review Report: HE-DeepFM: An FHE Inference System for CTR Prediction with Efficient FM Interactions

Academic fraud report · Geng Detector

Summary

Verdict: CLEAR (no fraud indicators detected). This SIGIR '26 conference paper by Su, Gu, Wang, Wang, Zheng, Cheng, Gong, and Wang proposes an optimized Fully Homomorphic Encryption (FHE) inference system for CTR prediction. Four checks were performed: (1) numerical self-consistency of Table 2 latency breakdowns — row sums match column totals within realistic floating-point rounding error (e.g., 2.5+1.4+7.9+49.0=60.8 vs reported 60.7; 2.9+13.9+8.5+147.7=173.0 vs 172.8); (2) timeline consistency — Intel Xeon Platinum 8488C and NVIDIA A100 hardware plus 2025 references ([13] Ebel et al., [15] Garimella et al.) align with a 2026 publication date; (3) mathematical validity — the FM O(n²)-to-O(1) reduction via the identity (∑uᵢ)² = ∑uᵢ² + 2∑uᵢuⱼ with CKKS SIMD packing and tree reduction is sound; (4) figure analysis limited to logical review, which found no duplication. Limitations: pixel-level image manipulation checks were not applicable. Confidence: high for the assessed dimensions; overall confidence moderate given the scope of automated checks.

Verdict

CLEAR. No evidence of data fabrication, image manipulation, mathematical inconsistency, or timeline conflict was identified. The paper passes the automated integrity checks performed.

Key findings

Data self-consistency (Table 2): Latency component sums match reported totals with small, irregular rounding differences characteristic of genuine empirical measurements (e.g., 60.8 vs 60.7; 173.0 vs 172.8; 560.6 vs 560.7). No signs of fabricated or synthetic data.
Timeline and environment consistency: Hardware (Intel Xeon Platinum 8488C / Sapphire Rapids, 2023; NVIDIA A100) and cited dependencies ([13] Ebel et al. 2025; [15] Garimella et al. 2025) are fully compatible with a 2026 publication at SIGIR '26 (July 20–24, 2026).
Mathematical derivation validity: Section 3.2 / Algorithm 1 correctly decomposes FM pairwise interactions using a sum-of-squares / square-of-sums identity and exploits CKKS SIMD batching with tree reduction, achieving the claimed complexity reduction.
Figure review (logical only): Figures 1–5 consist of architecture diagrams, algorithm flowcharts, and automatically generated plots; no internal numerical or legend inconsistencies were observed. Pixel-level duplication analysis was not applicable.

Evidence highlights

Table 2 row-sum checks (representative):
HE-DeepFM (HE-FM, Thres=500): 2.5 + 1.4 + 7.9 + 49.0 = 60.8 vs reported 60.7
HE-LRM (Baseline, Thres=5000): 2.9 + 13.9 + 8.5 + 147.7 = 173.0 vs reported 172.8
HE-DeepFM (Naive FM, Thres=50000): 5.8 + 490.2 + 9.0 + 55.6 = 560.6 vs reported 560.7
Identity used: $(\sum u_i)^2 = \sum u_i^2 + 2 \sum_{i<j} u_i \cdot u_j$, split into P (sum-then-square) and Q (square-then-sum) stages.
DOI: 10.1145/3805712.3809866

Notes

This is a conference paper in computer systems / applied cryptography; figures are primarily diagrams and code-generated plots, so standard biomedical image-forgery checks (e.g., Western blot, microscopy) do not apply.
Automated checks cannot fully replicate expert peer review; the absence of red flags here does not constitute a positive endorsement of novelty or significance.
No action recommended: original data not requested, no PubPeer comment suggested, no editorial notification warranted based on current evidence.