Towards Deep Encrypted Training: Low-Latency, Memory-Efficient, and High-Throughput Inference for Privacy-Preserving Neural Networks

2026-04-18

AI Summary

Current batched homomorphic encryption (HE)-based neural network inference suffers from low efficiency, hindering its applicability to high-throughput and training-oriented privacy-preserving computation. This work proposes a batch-oriented, HE-friendly neural network architecture together with a customized pipelined scheduling strategy, achieving for the first time highly efficient encrypted inference with high throughput, low latency, and reduced memory consumption. Leveraging an HE-friendly ResNet design, our approach reduces the amortized per-image inference time to 8.86 seconds when processing 512 encrypted images using ResNet-20, while cutting memory usage by a factor of 3.74. For ResNet-34 with a batch size of 256, the amortized inference time reaches 28.14 seconds, substantially overcoming the performance bottlenecks of existing methods in deep models and large-batch scenarios.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Security and Privacy: Privacy-enhancing technologies
• Responsible Web: Data and user privacy-enhancing technologies for the Web
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Privacy-preserving machine learning (PPML) has become increasingly important in applications where sensitive data must remain confidential. Homomorphic Encryption (HE) enables computation directly on encrypted data, allowing neural network inference without revealing raw inputs. While prior works have largely focused on inference over a single encrypted image, batch processing of encrypted inputs lags behind, despite being critical for high-throughput inference scenarios and training-oriented workloads. In this work, we address this gap by developing optimized algorithms for batched HE-friendly neural networks. We also introduced a pipeline architecture designed to maximize resource efficiency for different batch size execution. We implemented these algorithms and evaluated our work using HE-friendly ResNet-20 and ResNet-34 models on encrypted CIFAR-10 and CIFAR-100 datasets, respectively. For ResNet-20, our approach achieves an amortized inference time of 8.86 seconds per image when processing a batch of 512 encrypted images, with a peak memory usage of 98.96 GB. These results represent a 1.78x runtime improvement and a 3.74x reduction in memory usage compared to the state-of-the-art design. For the deeper ResNet-34 model, we achieve an amortized inference time of 28.14 on a batch of 256 encrypted images using 246.78GB of RAM

Problem

• Privacy-preserving machine learning
• Homomorphic Encryption
• Batch processing
• Encrypted inference
• High-throughput

Innovation

• Homomorphic Encryption
• Batched Inference
• Privacy-Preserving Machine Learning
• Memory-Efficient
• High-Throughput