Reproduce “Encrypt What Matters: When Selective Homomorphic Inference Is Efficient” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ali Backour, Juan Reyes, Jaime Punyed, Ana Onoprishvili
Selective homomorphic inference can substantially reduce encrypted computation when a model preserves spatial locality after a sensitive region is encrypted. The proposed dependency-propagation method exactly matches full-FHE outputs and achieves measured speedups of up to 30.63×, including 18.0× for VGG-11 on a 32×32 region, while predicting performance across architectures with a median error of 2.7%.
Key insights
- Selective-FHE efficiency is governed by how long a network maintains spatial locality, not simply by the encrypted input fraction, depth, or broad architecture category.
- For stride-one convolutions, an ROI grows by k−1 pixels per layer for a k×k kernel; once the tainted region saturates the feature map, the selective advantage largely disappears.
- Local architectures such as VGG-11 retain useful locality, whereas large-stride operations, global attention, dense layers, and global squeeze-and-excitation can make encrypted dependence nearly global early in the model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.