EI-DDLGN: Efficient Encrypted Inference with Deep Differentiable Logic Gate Networks under TFHE
Mahmoud Y. M. Yassin, Mahmoud AbdelHafeez Sayed, Mostafa Taha
cs.CR · cs.LG
Quick verdict
First TFHE study of logic-gate networks; 13.4x speedup over QAT-FCNN on MNIST is mainly driven by the Boolean-native architecture, not the proposed PBS bypass.
Summary
The paper explores Deep Differentiable Logic Gate Networks (DDLGNs) as a Boolean-native alternative to arithmetic neural networks for privacy-preserving inference under TFHE. Since discretized DDLGNs compute only two-input Boolean gates, encrypted inference maps directly onto TFHE's PBS-based Boolean operations, avoiding accumulator growth and bit-width sensitivity. The authors classify the 16 learned gate types into ten PBS-requiring and six PBS-free functions, and introduce Model-Fixed-Wire PBS Bypass, which detects wires that are constant for every possible input and substitutes them into gate truth tables, converting some PBS operations into forwarding, negation, or constants. Across 72 depth-width configurations (depth 1–6, width 2000–8000) on MNIST, FashionMNIST, and UCI Phishing, bypass rates average 5.3%, 3.9%, and 14.2%, respectively. On MNIST with TFHE-rs parameters (132-bit estimated security), EI-DDLGN-Small (2×4K) reaches 91.55% accuracy at 6.58 s/image versus 91.54% at 88.24 s for the reproduced QAT-FCNN-4 baseline — a 13.4× speedup; larger models reach 97.20% at 34.14 s. The authors honestly note that the speedup comes mainly from the Boolean-native structure, not the bypass. Limitations: external baselines are compared on MNIST only, DDLGN results lack seed averaging or variance reporting (the baseline uses 10 seeds), TFHE-oriented BNN baselines are discussed but not benchmarked, and the bypass technique's practical benefit is modest.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- DDLGNs are Boolean-native and therefore better aligned with TFHE's PBS-based Boolean execution model than QAT arithmetic NNs or BNNs.
- Encrypted execution cost of TFHE-based DDLGN inference depends jointly on model size, learned Boolean-function distribution, and propagated wire status, not gate count alone.
- MFW-PBS Bypass is a semantics-preserving execution strategy that eliminates part of the PBS budget (mean 3.9–14.2% per dataset) without modifying network topology.
- On MNIST, EI-DDLGN-Small matches QAT-FCNN-4 accuracy (91.55% vs 91.54%) at 13.4× lower latency (6.58 s vs 88.24 s per image).
- Larger EI-DDLGN models exceed all evaluated QAT-FCNN baselines in accuracy (up to 97.20%) while remaining 3.9× faster than the best-accuracy baseline.
Strengths
Genuinely novel positioning
First systematic study of DDLGN under TFHE (Sec. 1, Table 1): unlike QAT-FCNN or BNN adaptations that keep arithmetic accumulation, DDLGN neurons are learned two-input Boolean gates, so encrypted inference aligns natively with PBS-based Boolean execution.
Formal, semantics-preserving bypass
MFW-PBS Bypass (Sec. 5.2) is fully specified: wire status σ(w) ∈ {M0, M1, E} (Eq. 14), model-fixed status assigned only when a wire is constant over the whole input domain (Eq. 16), and Table 2 enumerates the truth-table substitution for all 10 PBS gates under all four fixed-input cases. The server never inspects or branches on plaintext data.
Systematic evaluation grid
Honest attribution of the speedup
Security-parameter diligence
Weaknesses
Thin baseline coverage
Asymmetric statistical treatment
Headline technique contributes little
Unquantified implementation confound in latency
Narrow scope and modest accuracy ceiling
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EI-DDLGN (arXiv:2609.13636, accepted at FPS 2026) proposes running Deep Differentiable Logic Gate Networks — networks whose neurons are learned two-input Boolean gates (Petersen et al., NeurIPS 2022) — under TFHE, replacing the dominant QAT-arithmetic paradigm. Four claims: (1) first in-depth study of DDLGN under TFHE; (2) encrypted cost depends jointly on model size, learned gate distribution, and propagated wire status rather than gate count alone; (3) MFW-PBS Bypass removes PBS operations semantics-preservingly without topology change; (4) on MNIST, EI-DDLGN-Small matches QAT-FCNN-4 accuracy (91.55% vs 91.54%) at 13.4× lower latency (6.58 s vs 88.24 s per image, Table 5). Scope is deliberately feedforward-only (CNNs deferred in Sec. 2), with an honest-but-curious threat model (Sec. 4.2) and no model confidentiality.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EI-DDLGN?
EI-DDLGN is a framework for running deep differentiable logic gate networks — neural networks whose neurons are learned two-input Boolean gates — as encrypted inference under TFHE fully homomorphic encryption, so a server can classify a client's data without ever seeing it. The paper is by Yassin, Sayed, and Taha (Carleton University), accepted at FPS 2026.
How much faster is EI-DDLGN than previous TFHE neural network inference?
On MNIST, the small model (2 layers × 4000 gates) reached 91.55% accuracy at 6.58 seconds per image, versus 88.24 seconds for an accuracy-matched quantization-aware FCNN baseline (91.54%) — about 13.4× faster. Larger variants reached 97.20% accuracy while still roughly 3.9× faster than the best-accuracy baseline. The authors note the speedup comes mainly from the Boolean-native architecture, not from their PBS-bypass optimization.
What is MFW-PBS Bypass?
Model-Fixed-Wire PBS Bypass is an execution optimization that identifies wires whose values are constant for every possible input, substitutes them into each gate's truth table, and thereby replaces some costly programmable bootstrapping (PBS) operations with cheap ciphertext forwarding, encrypted negation, or constants — without changing the trained network or inspecting plaintext data. Measured bypass rates averaged 5.3% on MNIST, 3.9% on FashionMNIST, and 14.2% on UCI Phishing.
Is the EI-DDLGN code and data available?
The paper states the implementation is available at github.com/Carleton-SCI/EI-DDLGN. Experiments use public datasets (MNIST, FashionMNIST, UCI Phishing), the difflogic training library, and the TFHE-rs Rust library with parameters estimated at 132-bit security, on a commodity Intel i9-10900 desktop.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–8 appear only as captions in the extracted text; plotted per-configuration values could not be verified, so figure-level judgments rely on the authors' in-text descriptions and on Tables 3–5.
- The linked GitHub repository was not inspected (no network access); code availability is reported as stated in the paper.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.