GPU CKKS path speeds encrypted MNIST training about 49 times
Original title: GPU-Accelerated Non-Interactive Encrypted Neural Network Training with Batched Diagonal Packing
In a matched 100-update run, device-resident bootstrapping cut training from 14,268.295 to 293.075 seconds with equal decoded parameters.
An MIT-licensed Rust CKKS implementation keeps encrypted neural-network training and bootstrap arithmetic on the GPU while host-side key generation and client operations stay on the CPU. On the main non-interactive MNIST model, a 100-update comparison showed a 48.68 times training-time ratio and a 40.67 times ratio including setup. A separate GPU run finished ten passes over 4,800 images in 2,967.095 seconds at 90.35% test accuracy. The authors do not claim unconditional superiority over ReBoot.
Unconfirmed: Full-setting ReBoot CPU time is unmeasured, and the baseline-batch normalized key-switch target was missed.
- Decoded parameters and predictions matched the CPU run at 100 updates.
- GPU test accuracy was 90.35% versus 90.41% for the plaintext twin.
- Exact CPU and GPU agreement was not verified for the full 1,000-update run.
- The released ReBoot baseline misses the common security-estimator threshold.
From the original article
1 min read at arxiv.orgTraining a neural network under homomorphic encryption requires transposes, batch reductions, optimizer updates, and repeated refreshes in addition to forward matrix products. We combine rectangular diagonal packing, feature-major SIMD mini-batch lanes, baby-step giant-step evaluation, and compact optimizer state in an MIT-licensed Rust CKKS implementation.
By Kentaro Mihara · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.