Skip to content

Private and reliable neural network inference

Authors: Nikola Jovanovic; Marc Fischer; Samuel Steffen; Martin Vechev
Year: 2022
Venue: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security

type: system collection: sok-references status: auto

Links: DOI · arXiv:2210.15614

Summary

Auto-extracted summary (unreviewed)

Phoenix is the first system combining privacy-preserving inference with reliability guarantees: it designs efficient RNS-CKKS counterparts (e.g. homomorphic sign and argmax, soft preliminary counting) for the control-flow-heavy building blocks of randomized smoothing, enabling encrypted inference with robustness and fairness certificates. Implemented on SEAL 3.6.6 at 128-bit security (N=2^17) and evaluated on small networks for MNIST, Adult and CIFAR-10 on a 64-core Xeon machine, Phoenix produces results almost always consistent with the non-private procedure without prohibitive latency.

Abstract

Reliable neural networks (NNs) provide important inference-time reliability guarantees such as fairness and robustness. Complementarily, privacy-preserving NN inference protects the privacy of client data. So far these two emerging areas have been largely disconnected, yet their combination will be increasingly important. In this work, we present the first system which enables privacy-preserving inference on reliable NNs. Our key idea is to design efficient fully homomorphic encryption (FHE) counterparts for the core algorithmic building blocks of randomized smoothing, a state-of-the-art technique for obtaining reliable models. The lack of required control flow in FHE makes this a demanding task, as naïve solutions lead to unacceptable runtime. We employ these building blocks to enable privacy-preserving NN inference with robustness and fairness guarantees in a system called Phoenix. Experimentally, we demonstrate that Phoenix achieves its goals without incurring prohibitive latencies. To our knowledge, this is the first work which bridges the areas of client data privacy and reliability guarantees for NNs.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution FHE building blocks for randomized smoothing (sign, argmax, counting) without control flow llm 0.85
contribution Phoenix system for private inference with robustness and fairness guarantees llm 0.85
contribution Analysis of inconsistency sources (range/difference violations, CKKS noise, soft counting) with error-probability bounds llm 0.85
contribution MockSEAL cleartext simulator for large-scale testing llm 0.85
schemes CKKS llm 0.85
libraries SEAL 3.6.6 llm 0.85
libraries MockSEAL llm 0.85
models 2-layer fully-connected network (hidden 32) llm 0.85
models 3-layer network (Juvekar et al. network A) llm 0.85
datasets MNIST llm 0.85
datasets Adult llm 0.85
datasets CIFAR-10 llm 0.85
tasks private NN inference llm 0.85
tasks certified robustness via randomized smoothing llm 0.85
tasks fairness certification llm 0.85
hardware 2-socket Intel Xeon Gold 6242 2.80GHz, 64 cores llm 0.85
security level 128-bit llm 0.85
threat model client-server MLaaS; client data hidden from server llm 0.85
interaction non-interactive FHE llm 0.85
code url https://github.com/eth-sri/phoenix (URL found in the full text) regex 0.50
code url https://github.com/eth-sri/phoenix llm 0.85
limitations FHE primitives are orders of magnitude slower than cleartext, so extensive testing in real FHE is impractical (MockSEAL used instead) llm 0.85
limitations Harmful range violations of the sign approximation can compromise an entire inference llm 0.85
limitations Soft preliminary counting heuristic can produce results inconsistent with the non-private procedure llm 0.85
llm relevance medium llm 0.85
key result consistency with non-private evaluation: inconsistencies almost nonexistent (MNIST, Adult, CIFAR-10 with randomized smoothing, N=2^17, scale 2^50) llm 0.85

SoK reference [81]

N. Jovanovic, M. Fischer, S. Steffen, and M. T. Vechev, “Private and reliable neural network inference,” in ACM CCS 2022: 29th Conference on Computer and Communications Security, H. Yin, A. Stavrou, C. Cremers, and E. Shi, Eds. Los Angeles, CA, USA: ACM Press, Nov. 7–11, 2022, pp. 1663–1677.

Listed on the reference page of the site.

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: CKKS 34 [p.2], MPC 5 [p.1], BFV 3 [p.13], BGV 3 [p.13], TFHE 2 [p.13], Scheme switching 2 [p.14], Leveled / somewhat HE 2 [p.4]
  • library: Microsoft SEAL 14 [p.2], EVA / CHET 5 [p.11], TenSEAL 2 [p.14], cuHE 1 [p.14]
  • model: CNN 3 [p.14], AlexNet 1 [p.14]
  • dataset: MNIST 16 [p.11], Adult (census income) 2 [p.11], CIFAR-10 1 [p.11]
  • hardware: GPU 2 [p.13], Intel Xeon 1 [p.10]
  • technique: RNS 31 [p.2], Bootstrapping 9 [p.4], ReLU 7 [p.5], Rotations 5 [p.4], SIMD packing 4 [p.4], Compiler 4 [p.14], Rescaling 3 [p.8], Convolution 3 [p.15], Non-interactive 3 [p.12], Semi-honest model 3 [p.3], Model extraction / stealing 3 [p.4]
  • assumption: LWE 2 [p.9]
  • task: Medical / healthcare 19 [p.1], Finance 6 [p.1]
  • topic: Homomorphic encryption 95 [p.1], Private / secure inference 7 [p.5], Open source 5 [p.2], Privacy-preserving machine learning 4 [p.2]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/sok-references/081_private-and-reliable-neural-network-inference.pdf sok-references 18 1.0 MB ba96efdb7d95

Database key: private-and-reliable-neural-network-inference · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set private-and-reliable-neural-network-inference COLUMN VALUE