Research map / Federated learning and privacy / Privacy attacks and secure learning
PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PrivDNFIS, a lattice-based (RLWE/LWE somewhat homomorphic encryption) privacy-preserving inference framework for deep neuro-fuzzy inference systems that avoids ciphertext rotations and replaces softmax with a quadratic function.
Problem
Privacy-preserving inference over deep neuro-fuzzy inference systems (DNFIS), where users' sensitive inputs must stay hidden from cloud servers, has not been studied before.
Approach
Packs feature vectors into polynomial coefficients so inner products and sums are read from polynomial coefficients via ciphertext extraction (no rotations), computes logarithmic Gaussian memberships with quadratic polynomials, and replaces the output softmax with a quadratic approximation whose ranking is unchanged.
Key results
- PrivDNFIS reduces end-to-end time cost roughly 1.9x on CIFAR-10 and 4.4x on CIFAR-100 versus the CryptFlow benchmark.
- Processing 200 queries on CIFAR-100 takes 1194.62 s in total (about 6 s per query).
- Inference accuracy is unchanged from the non-private DCNFIS: 93.02% on CIFAR-10 and 74.52% on CIFAR-100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace softmax with a quadratic function (ablation) | PrivDNFIS · CIFAR-10 | accuracy: 93.02% -> 93.02% | No clear effect |
| replace softmax with a quadratic function (ablation) | PrivDNFIS · CIFAR-100 | accuracy: 74.52% -> 74.52% | No clear effect |
| use coefficient encoding and extraction to compute vector inner products without ciphertext rotation | PrivDNFIS · CIFAR-10 | end-to-end speedup: 1.9× | Helped |
| use coefficient encoding and extraction to compute vector inner products without ciphertext rotation | PrivDNFIS · CIFAR-100 | end-to-end speedup: 4.4× | Helped |
| replace softmax with a quadratic function (ablation) | PrivDNFIS · CIFAR-10 and CIFAR-100 | inference efficiency | Helped |
Surprising findings
- Replacing softmax with a quadratic function is claimed to incur no accuracy loss because the score ranking is preserved.
Limitations
- Security is under the semi-honest model only; malicious users are left to future work.
- Full security proof is deferred to a full version; only a proof sketch is given.
- Returns top-k scores and offloads reciprocal computation to the user, which reveals intermediate values to the user.
Experimental setup
Datasets: CIFAR-10 CIFAR-100
Benchmarks: CIFAR-10 CIFAR-100
Metrics: end-to-end running time encryption time communication cost decryption time classification accuracy
Baselines: CryptFlow
Builds on: DNFIS (Yeganejou et al. 2023) ANFIS (Jang 1993) BFV Chen et al. 2021 ciphertext extraction CryptFlow
Closest papers by meaning
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.86)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · similarity 0.84)
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.83)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.83)
- HEPrune: Fast Private Training of Deep Neural Networks With Encrypted Data Pruning (NeurIPS 2024 · similarity 0.82)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.