Paper 2018/1080

How Does Strict Parallelism Affect Security? A Case Study on the Side-Channel Attacks against GPU-based Bitsliced AES Implementation

Yiwen Gao, Yongbin Zhou, and Wei Cheng

Abstract

Parallel cryptographic implementations are generally considered to be more advantageous than their non-parallel counterparts in mitigating side-channel attacks because of their higher noise-level. So far as we know, the side-channel security of GPU-based cryptographic implementations have been studied in recent years, and those implementations then turn out to be susceptible to some side-channel attacks. Unfortunately, the target parallel implementations in their work do not achieve strict parallelism because of the occurrence of cached memory accesses or the use of conditional branches, so how strict parallelism affects the side-channel security of cryptographic implementations is still an open problem. In this work, we make a case study of the side-channel security of a GPU-based bitsliced AES implementation in terms of bit-level parallelism and thread-level parallelism in order to show the way that works to reduce the side-channel security of strict parallel implementations. We present GPU-based bitsliced AES implementation as the study case because (1) it achieves strict parallelism so as to be resistant to cache-based attacks and timing attacks; and (2) it achieves both bit-level parallelism and thread-level parallelism (a.k.a. task-level parallelism), which enables us to research from multiple perspectives. More specifically, we first set up our testbed and collect electro-magnetic (EM) traces with some special techniques. Then, the measured traces are analyzed in two granularity. In bit-level parallelism, we give a non-profiled leakage detection test before mounting attacks with our proposed bit-level fusion techniques like multi-bits feature-level fusion attacks (MBFFA) and multi-bits decision-level fusion attacks (MBDFA). In thread-level parallelism, a profiled leakage detection test is employed to extract some special information from multi-threads leakages, and with the help of those information our proposed multi-threads hybrid fusion attack (MTHFA) method takes effect. Last, we propose a simple metric to quantify the side-channel security of parallel cryptographic implementations. Our research shows that the secret key of our target implementation can be recovered with less cost than expected, which suggests that the side-channel security of parallel cryptographic implementations should be reevaluated before application.