Paper 2022/1472

Hardware-Supported Cryptographic Protection of Random Access Memory

Abstract

Confidential Computing is the protection of data in use from access or modification by any unauthorized agent, including privileged software. For example, in Intel SGX (Client and Scalable versions) and TDX, AMD SEV, Arm CCA, and IBM Ultravisor this protection is implemented via access control policies. Some of these architectures also include memory protection schemes relying on cryptography, to protect against physical attacks. We review and classify such schemes, from academia and industry, according to protection levels corresponding of adversaries with varying capabilities, budget, and strategy. The building blocks of all memory protection schemes are encryption and integrity primitives and modes of operation, as well as anti-replay structures. We review these building blocks, consider their possible combinations, and evaluate the performance impact of the resulting schemes. We present a framework for performance evaluation in a simulated system. To understand the best and worst case overhead, systems with varying load levels are considered. We propose new solutions to further reduce the performance and memory overheads of such technologies. Advanced counter compression techniques make it viable to store counters used for replay protection in a physically protected memory. By additionally repurposing some ECC bits to store integrity tags, we can provide the highest levels of confidentiality, integrity, and replay protection at a hitherto unattained performance penalty, namely 3.32%, even under extreme load and at costs that make them reasonable in data centers. Combinations of technologies that are suitable for client devices are also discussed.

Note: Discussed more results/observations.