Characterization and Methods to Improve the Efficiency of Machine learning under Homomorphic Encryption

Dr. Subhankar Pal - IBM

Nov. 24, 2023, 10:45 a.m. - Nov. 24, 2023, 11:30 a.m.

McConnell Eng. Room 603

Hosted by: Christophe Dubach

Fully Homomorphic Encryption (FHE) is a cryptographic paradigm that enables secure computation with end-to-end quantum-safe encryption. However, FHE computations come with substantial performance and energy overheads compared to the equivalent unencrypted operations on plaintext. Several algorithmic improvements have reduced this overhead, but encrypted computations remain 3-4 orders of magnitude slower and less efficient than the corresponding unencrypted operations. We first present a detailed characterization of popular FHE libraries on a server-class CPU and GPU across key FHE primitives and a Privacy-Preserving Machine Learning (PPML) application. We then propose a framework called HE-PEx that uses pruning on top of a packing technique called tile tensors to reduce the latency and memory of PPML inference. HE-PEx uses permutations to prune additional ciphertexts and expansion to recover inference loss. We demonstrate the effectiveness of our methods for pruning fully-connected and convolutional layers in neural networks (NNs) on various PPML tasks. We implement and deploy networks atop a framework called HElayers, which shows a 10-35% improvement in inference speed and a 17-35% decrease in memory requirement over the unpruned network, within a 2.5% degradation in inference accuracy over the unpruned network.

Dr. Subhankar Pal is a Research Scientist in the Efficient and Resilient Systems group at IBM T.J. Watson Research Center, NY, USA. He is broadly interested in computer architecture and hardware-software co-design. He presently works on accelerating homomorphic encryption and early-stage design space exploration of systems-on-chips.