Skip to content

Enhancing privacy-preserving machine learning with self-learnable activation functions in fully homomorphic encryption

Authors: Jichao Xiong; Jiageng Chen; Junyu Lin; Dian Jiao; Hui Liu
Year: 2024
Venue: Journal of Information Security and Applications

type: technique collection: ppml status: auto

Links: DOI

Summary

Auto-extracted summary (unreviewed)

Proposes a Self-Learnable Activation Function (SLAF) and restructured linear layers so that CNNs can be made deeper under fully homomorphic encryption without large computational overhead, targeting privacy-preserving biometric (face) authentication. On the UTKFace dataset the SLAF-based model improves accuracy by 0.88-3.15% over networks with simple polynomial activations and by 4.87-9.67% over CryptoNets, and a lighter SNN variant responds in 0.778 s at some accuracy cost.

Abstract

In the field of artificial intelligence and data engineering, the effective utilization of data is critical for improving productivity across various sectors. As machine learning increasingly relies on sensitive data, balancing privacy with computational efficiency has become a major challenge. Homomorphic encryption provides a promising solution by enabling computation on encrypted data while preserving privacy in machine learning. However, its integration with neural networks is hindered by high computational demands and limitations in performing complex nonlinear operations. To address these challenges, we propose a novel approach that incorporates a "Self-Learnable Activation Function" (SLAF) and refines the structure of neural network linear layers. These enhancements are designed to accommodate the constraints of homomorphic encryption, allowing for deeper network architectures without significant computational overhead. Our optimized neural network model, tailored for biometric authentication tasks, outperforms traditional methods that use simple polynomial activation functions. Using the UTKFace dataset, which includes facial features under diverse scenarios, we validated the effectiveness of our solution in real-world applications. Experimental results demonstrate accuracy improvements of 0.88% to 3.15% over traditional models and 4.87% to 9.67% over the CryptoNets model, underscoring the capability of our approach to meet stringent privacy-preserving biometric authentication requirements.
Source: pdf

Extracted details

Field Value Source Confidence
contribution Self-Learnable Activation Function (SLAF) for HE-friendly neural networks llm 0.75
contribution Refined linear-layer structure allowing deeper networks under FHE llm 0.75
contribution Two models (SLAFNN for accuracy, SNN for speed) for privacy-preserving biometric authentication llm 0.75
contribution Evaluation on UTKFace against polynomial-activation baselines and CryptoNets llm 0.75
libraries SEAL llm 0.75
models CNN (SLAFNN) llm 0.75
models SNN llm 0.75
models CryptoNets (baseline) llm 0.75
datasets UTKFace llm 0.75
tasks biometric authentication llm 0.75
tasks face image classification llm 0.75
security level 128-bit llm 0.75
interaction non-interactive FHE llm 0.75
code url https://github.com/Microsoft/SEAL.Microsoft (URL found in the full text) regex 0.50
limitations The faster SNN model shows some accuracy degradation relative to SLAFNN llm 0.75
llm relevance medium llm 0.75
key result accuracy improvement: +0.88% to +3.15% (vs traditional polynomial-activation models, UTKFace) llm 0.75
key result accuracy improvement: +4.87% to +9.67% (vs CryptoNets, UTKFace) llm 0.75
key result response time: 0.778 s (SNN fast model under FHE) llm 0.75

Keyword mentions

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

  • scheme: CKKS 15 [p.2], Scheme switching 5 [p.2], TFHE 4 [p.2], MPC 4 [p.2], BGV 3 [p.2], FHEW 3 [p.2], YASHE 2 [p.8]
  • library: Microsoft SEAL 4 [p.3], EVA / CHET 3 [p.2], TenSEAL 2 [p.3]
  • model: Spiking neural network 22 [p.6], CNN 15 [p.1], ResNet 1 [p.2], GNN 1 [p.10], Autoencoder 1 [p.10], Mixture of experts 1 [p.10]
  • dataset: MNIST 11 [p.2], Fingerprint (FVC) 3 [p.1], CIFAR-10 1 [p.2]
  • hardware: GPU 3 [p.2], Apple silicon 3 [p.3]
  • technique: Convolution 31 [p.1], ReLU 22 [p.2], Square activation 10 [p.2], Bootstrapping 6 [p.2], Federated learning 5 [p.2], Matrix multiplication 3 [p.4], Rescaling 2 [p.3], Polynomial approximation 2 [p.2], Lookup table 2 [p.2]
  • assumption: RLWE 4 [p.2]
  • task: Image classification 3 [p.5], Medical / healthcare 2 [p.10]
  • topic: Homomorphic encryption 93 [p.1], Privacy-preserving machine learning 9 [p.1], Benchmark 4 [p.2]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/inference/xiong-2024_enhancing-ppml-with-self-learnable-activation-functions-in-fhe.pdf ppml/inference 10 986.1 KB 1498c65b7557 was Xiong et al. - 2024 - Enhancing privacy-preserving machine learning with self-learnable activation functions in fully homo.pdf

Database key: enhancing-privacy-preserving-machine-learning-with-self-learnable · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set enhancing-privacy-preserving-machine-learning-with-self-learnable COLUMN VALUE