Enabling Homomorphically Encrypted Inference for Large DNN Models

Abstract

The proliferation of machine learning services in the last few years has raised data privacy concerns. Homomorphic encryption (HE) enables inference using encrypted data but it incurs 100x-10,000x memory and runtime overheads. Secure deep neural network (DNN) inference using HE is currently limited by computing and memory resources, with frameworks requiring hundreds of gigabytes of DRAM to evaluate small models. To overcome these limitations, in this paper we explore the feasibility of leveraging hybrid memory systems comprised of DRAM and persistent memory. In particular, we explore the recently-released Intel R Optane TM PMem technology and the Intel R HE-Transformer nGraph R to run large neural networks such as MobileNetV2 (in its largest variant) and ResNet-50 for the first time in the literature. We present an in-depth analysis of the efficiency of the executions with different hardware and software configurations. Our results conclude that DNN inference using HE incurs on friendly access patterns for this memory configuration, yielding efficient executions.

ecoHMEM: Improving Object Placement Methodology for Hybrid Memory Systems in HPC

“…The experiments were performed in a machine equipped with 2 Intel ® Xeon ® Platinum 8260L processors running at a nominal frequency of 2.3GHz, 4×8 GB DDR4 DIMMs and 12×512 GB Intel ® Optane ™ PMem 100 series DIMMs running at 2666 MT/s. The ratio of DRAM to PMem is intentionally lower than the recommended by the vendor, advocating for more energy friendly and cost-efficient system configurations [23]; results with only 4 PMem DIMMs are also presented to assess the impact of this ratio. The server runs Linux Fedora 27 (kernel 4.18.8) and we used the Intel compiler and Intel MPI version 2019.5.281.…”

Outsourced Databases in the Cloud: A Privacy-Preserving Indexing Scheme

“…Ideally, robust encryption schemes with strong security guarantees, such as fully homomorphic encryption [? ], [44], [24], oblivious RAM [26], or searchable encryption [17] are the best solutions. However, current schemes with strong security levels typically require DBMS changes, which is not always feasible in cloud outsourcing scenarios.…”