GhostLeg: Selective Memory Coalescing for Secure GPU Architecture

Architectural considerations for secure executions are getting more critical GPU since popular security applications and libraries have been ported to a domain rely on GPU’s massively parallel computations. Recent studies disclosed the attack models that exploit architectural vulnerabilities leak secret keys of AES. The high correlations between execution time kernel number memory requests generated from coalescing. Thus prior defenses provide by randomizing or restricting coalescing load warps. However, those defense approaches result in significant performance overhead is an essential feature improving GPU. In this paper, we propose GhostLeg, efficient approach against correlation-based attacks. GhostLeg selectively applies warps minimize induced concealing behavior. Our analysis AES reveals only whose index addresses influenced vulnerable order executions, pinpoints require based class source register. flag assigned each register can be set propagation non-deterministic user data (GhostLeg-ND) specific directive marked programmers (GhostLeg-Key). evaluation shows guarantees attacks GhostLeg-ND exhibits 54.7% higher compared state-of-the-art solution.