Low-cost mechanisms for isolated cryptographic operations on commodity accelerators

Abstract

General-Purpose computing on a Graphics Processing Unit (GPGPU) involves leveraging commodity GPUs as massively parallel processing units. GPGPU is an emerging computing paradigm for high-performance and data-intensive computations such as cryptographic operations. Although GPGPU is an attractive solution for accelerating modern cryptographic operations, the security challenges that stem from utilizing commodity GPUs remain an unresolved problem. In this dissertation, we investigate the protection of a cryptographic key during the acceleration of cryptographic operations on commodity GPUs under a broad range of GPU vulnerabilities and threats to the underlying host system. We show that the GPU cache can be used as a securely isolated place to store security-sensitive code and secret keys, where access is denied even to attackers with kernel privileges. Moreover, by carefully manipulating the GPU memory hierarchy, part of the GPU cache can be used exclusively for the secret key. Second, we present an On-demand Bootstrapping Mechanism for Isolated cryptographic operations (OBMI). By leveraging System Management Mode (SMM), a privileged execution mode provided by x86 architectures, OBMI implements a program and a secret key into the GPU such that they are securely isolated during the acceleration of cryptographic operations, even in the presence of compromised kernels. Our approach does not require an additional hardware-abstraction layer, and it does not entail modifying the GPU driver. An evaluation of the proposed OBMI demonstrated that even adversaries with kernel privileges cannot gain access to the secret key, and it also showed that the proposed mechanism incurs negligible performance degradation for both the CPU and GPU. Finally, we suggest the utilization of SMM which enables the isolated cryptographic operations on commodity x86 processor. We show how SMM can achieve the efficient cryptographic operations, moreover, how it can make synergy with suggested GPGPU-based isolated cryptographic operations.