Are Randomized Caches Truly Random? Formal Analysis of Randomized-Partitioned Caches

Cache based side-channel attacks exploit the fact that an adversary can setup the shared cache memory (the last level cache in modern systems) into a known state and detect any microarchitectural state changes made by the victim on the cache. Different mitigation techniques have been proposed in the literature that aims to mitigate these attacks by randomizing the address to cache location mappings. The security guarantees in these schemes are based on the degree of difficulty for an attacker to reliably determine the cache lines accessed by the victim within practical time settings. However, prior attacks have shown that newer and more improved algorithms can be envisaged that discover conflicting sets in the secured randomized caches. In this work, we first categorize different types of cache designs into four broad classes based on the extent of non-determinism and randomness of allocating an address in those caches. We then develop a mathematical framework to formally analyse the security implications of the randomized and partitioned cache designs in terms of collision probability, self-collision probability and size of the eviction set required to perform a successful eviction-based attack. We further empirically demonstrate set associative eviction on recently proposed randomization schemes called Mirage and Scattercache. Next, we propose two algorithms to generate efficient eviction set on these schemes and analytically evaluate the efficacy of our algorithms against the one proposed in the literature. Finally, we argue that mere randomization using a cryptographic primitive as used in popular schemes like Scattercache, CEASER-S, Mirage etc. does not provide the required randomness. Although the randomized-partitioned caches provide some resilience against eviction-set generation techniques, they are still vulnerable to eviction-based attacks.