PhantomCache: Obfuscating Cache Conflicts with Localized Randomization

—Cache conflicts due to deterministic memory-to-cache mapping have long been exploited to leak sensitive information such as secret keys. While randomized mapping is fully investigated for L1 caches, it still remains unresolved about how to secure a much larger last-level cache (LLC). Recent solutions periodically change the mapping strategy to disrupt the crafting of conflicted addresses, which is a critical attack procedure to exploit cache conflicts. Remapping, however, increases both miss rate and access latency. We present PhantomCache for securing an LLC with remapping-free randomized mapping. We propose a localized randomization technique to bound randomized mapping of a memory address within only a limited number of cache sets. The small randomization space offers fast set search over an LLC in a memory access. The intrinsic randomness still suffices to obfuscate conflicts and disrupt efficient exploitation of con-flicted addresses. We evaluate PhantomCache against an attacker exploring the state-of-the-art attack with linear-complexity. To secure an 8-bank 16 MB 16-way LLC, PhantomCache confines randomization space of an address within 8 sets and brings only 1.20% performance degradation on individual benchmarks, 0.50% performance degradation on mixed workloads, and 0.50% storage overhead per cache line, which are 2x and 9x more effi-cient than the state-of-the-art solutions. Moreover, PhantomCache is solely an architectural solution and requires no software change.