RECAST: Mitigating Conflict-Based Cache Attacks Through Fine-Grained Dynamic Mapping

Conflict-based cache attacks can leak critical information from target programs. Accordingly, randomization-based cache designs have emerged as an efficient and LLC-favorable way to mitigate such attacks. However, later investigations have revealed several problems with these designs. Specifically, we identify limited randomness and coarse-grained protection as key issues of previous designs. To solve these issues, we propose Recast, a secure cache design with address-sensitive secret generation and tweakable index randomization. Our insight is that cache modules at different levels can work collaboratively to enhance their security. Address-sensitive secret generation in private caches generates a secret value for each address upon cache misses. The shared cache in Recast uses tweakable index randomization, where the cryptographic function uses the secret value from private caches as the input to calculate the cache set index. Therefore, Recast achieves fine-grained dynamic mapping. We implement Recast in the gem5 simulator. We use a micro-benchmark and a benchmark suite to showcase the security of Recast. Our performance evaluations on SPEC 2017 and PARSEC benchmarks show that Recast incurs 2.29% and 2.03% performance overhead. Moreover, Recast with the LRU replacement policy has only 0.51% and 1.04% performance overhead on the two benchmarks. Therefore, Recast provides higher security guarantees with minimal performance overhead.