Reproduce “Memory-Efficient Designs for Word-Wise Universal Fully Homomorphic Encryption” 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
Ardhi Wiratama Baskara Yudha, Erwin Eko Wahyudi, Rian Adam Rajagede, Qian Lou, Yan Solihin
The BXT framework introduces memory-efficient designs for word-wise Universal Fully Homomorphic Encryption (uFHE) on GPUs, mitigating the memory bottleneck by integrating ciphertext compression, serialization, and operation-level optimizations. It achieved up to 5.8x speedup for CNN workloads and 9.2x speedup when combined with GME, while fault-aware training maintained accuracy despite significant precision reduction.
Key insights
- Memory efficiency, rather than solely computational speed, is critical for advancing FHE performance on GPUs due to inherent ciphertext expansion and low arithmetic intensity.
- Domain-specific optimizations like PRNG-based compression and bit-array serialization are effective for FHE ciphertexts, where generic compression methods are largely ineffective.
- Application-level pruning of ciphertext digits for comparison operations, when combined with fault-aware training, allows for significant performance gains with manageable accuracy loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.