Invalidate+Compare: A Timer-Free GPU Cache Attack Primitive

While extensive research has been conducted on CPU cache side-channel attacks, the landscape of similar studies on modern GPUs remains largely uncharted. In this paper, we investigate potential information leakage threats posed by the caches in GPUs of NVIDIA's latest Ampere and Ada Lovelace generations. We first exploit a GPU cache maintenance instruction to reverse engineer certain key properties of the cache hierarchy in these GPUs, and then we introduce a novel GPU cache side-channel attack primitive named INVALIDATE+COMPARE that is designed to spy on the GPU cache activities of a victim in a timer-free manner. We further showcase the use of this primitive with two case studies. The first one is a website fingerprinting attack that can accurately identify the web pages visited by a user, while the second one uncovers keystroke data entered via a virtual keyboard. To our knowledge, these stand as the first demonstrations of timer-free cache sidechannel attacks on GPUs.