DeepReShape: Redesigning Neural Networks for Efficient Private Inference

Published: 17 Jul 2024, Last Modified: 17 Sept 2024Accepted by TMLREveryoneRevisionsBibTeXCC BY 4.0
Abstract: Prior work on Private Inference (PI)---inferences performed directly on encrypted input---has focused on minimizing a network's ReLUs, which have been assumed to dominate PI latency rather than FLOPs. Recent work has shown that FLOPs for PI can no longer be ignored and incur high latency penalties. In this paper, we develop DeepReShape, a technique that optimizes neural network architectures under PI's constraints, optimizing for both ReLUs {\em and} FLOPs for the first time. The key insight is strategically allocating channels to position the network's ReLUs in order of their criticality to network accuracy, simultaneously optimizes ReLU and FLOPs efficiency. DeepReShape automates network development with an efficient process, and we call generated networks HybReNets. We evaluate DeepReShape using standard PI benchmarks and demonstrate a 2.1% accuracy gain with a 5.2$\times$ runtime improvement at iso-ReLU on CIFAR-100 and an 8.7$\times$ runtime improvement at iso-accuracy on TinyImageNet. Furthermore, we investigate the significance of network selection in prior ReLU optimizations and shed light on the key network attributes for superior PI performance.
Submission Length: Long submission (more than 12 pages of main content)
Changes Since Last Submission: Following are the changes made in the revised draft after the reviewers' comments. 1. The background section now includes one paragraph discussing the different protocols (GC vs OT) used for non-linear layers in private inference. We primarily discussed the recent work, CoPriv (NeurIPS'23). 2. We've rewritten the Discussion section (Section 7) to include broader impact, limitations, and future work. 3. We've re-organized the Appendix.
Video: https://drive.google.com/file/d/17byDYNIuiAYvdTQJJDv29UnK7XdCdwHQ/view?usp=sharing
Code: https://github.com/Nandan91/deepreshape-release
Assigned Action Editor: ~Caglar_Gulcehre1
Submission Number: 1732
Loading