Key recovery attack and analysis on GPUs

Abstract

In addition to graphic operations, graphics cards are used as computing devices for various purposes such as physical system simulation and machine learning. Recently, encryption algorithms have been offloaded to the graphics card, which shows a speedup of up to 28 times. In this situation, an attack to recover the key using the time difference between AES calculations on the graphics card was introduced in 2016. In this thesis, we apply the above attacks on the graphics cards of Maxwell and Pascal generations, and show that the attack does not work. The signal of the key recovery attack is defined as SIMT vulnerabilities, and the noise is defined as timing variance. Maxwell and Pascal generation graphics cards show low signal and high noise. The reason for the lower SIMT vulnerability is shown to be higher interconnect channel bandwidth. The reason for the increased timing variance is that there are periodic timing jumps when read from the constant memory region. We also tried to increase the signal level by saturating the interconnect channel bandwidth, and trying to find the timing channel in the noise through the percentage filter.