AccessShadow: Leveraging Adversarial Samples to Counter Deep Learning-Assisted Cache Timing Attacks

Cache timing attacks have remained a significant concern for two decades within the domain of side-channel attacks. They have effectively compromised confidentiality and breached privacy in various scenarios. With the integration of deep learning, the potency of cache timing side-channel attacks has escalated. However, traditional defensive measures may not adequately countermeasure against these deep learning-aided attacks, given that deep learning models have the ability to distill complex patterns and understandings from raw data. The models can detect and decode associations which are not discernible via traditional statistical methodologies, making it possible for them to exploit even subtle timing discrepancies and cache usage patterns, something often missed by conventional defenses. To counteract deep learning-assisted cache timing attacks, this work develops a software-based defense strategy, leveraging the concept of adversarial examples. By inserting a number of false memory accesses to the victim code without affecting the functionality, the timing traces gathered by the cache monitor contain deliberate noise, which effectively obscures valuable information leakage and reduces the susceptibility to the follow-on deep learning-assisted attacks. Our defense measure has demonstrated consistent resilience across a variety of cache types and has been effective on Intel X86 processors. The minimal overhead of this method represents a promising step towards practical, costefficient defense against increasingly sophisticated cache attacks.