Prime+Scope: Overcoming the Observer Effect for High-Precision Cache Contention Attacks

Modern processors expose software to information leakage through shared microarchitectural state. One of the most severe leakage channels is cache contention, exploited by attacks referred to as PRIME+PROBE, which can infer fine-grained memory access patterns while placing only limited assumptions on attacker capabilities. In this work, we strengthen the cache contention channel with a near-optimal time resolution. We propose PRIME+SCOPE, a cross-core cache contention attack that performs back-to-back cache contention measurements that access only a single cache line. It offers a time resolution of around 70 cycles (25ns), while maintaining the wide applicability of PRIME+PROBE. To enable such a rapid measurement, we rely on the deterministic nature of modern replacement policies and their (non-)interaction across cache levels. We provide a methodology to, essentially, prepare multiple cache levels simultaneously, and apply it to Intel processors with both inclusive and non-inclusive cache hierarchies. We characterize the resolution of PRIME+SCOPE, and confirm it with a cross-core covert channel (capacity up to 3.5 Mbps, no shared memory) and an improved attack on AES T-tables. Finally, we use the properties underlying PRIME+SCOPE to bootstrap the construction of the eviction sets needed for the attack. The resulting routine outperforms state-of-the-art techniques by two orders of magnitude. Ultimately, our work shows that interference through cache contention can provide richer temporal precision than state-of-the-art attacks that directly interact with monitored memory addresses.