FEnc$^2$: Unifying Data Packing for Efficient Private Inference via Convolution and Architecture-Aware Fragment Encoding
Researchers have developed FEnc$^2$, a new framework designed to significantly improve the efficiency of private inference using Fully Homomorphic Encryption (FHE). This method unifies data packing by considering both convolutional operations and network architecture to optimize ciphertext utilization and reduce computational overhead. FEnc$^2$ achieves substantial speedups, with end-to-end latency reductions of up to 228x on GPUs and 226x on CPUs for common image recognition tasks. AI
IMPACT Optimizes FHE for private ML, potentially enabling wider adoption of privacy-preserving AI applications.