Roland Czerny
← back

Fast and Secure LLC Caches via Spatial Windows

2026-07-01 · DIMVA 2026
Roland Czerny, Simon Lammer, Lukas Giner, Anirban Chakraborty, Daniel Gruss

Abstract

We propose a generic technique, ObfusCache, to de-correlate cache activity from secret-dependent operations in secure randomized last-level caches (LLCs) with configurable spatial windows. We reduce the side-channel-observable spatial granularity to the order of magnitude of TLB or page cache channels. We show that the randomization window acts as a prefetcher, improving performance by up to 8% in the Scimark2 benchmark compared to ScatterCache, which has no prefetching for security reasons. Regarding security, we show that the lower bound for the slowdown for Prime+Prune+Probe eviction set generation is at least three orders of magnitude (680x). More importantly, the required eviction set is larger than the cache, which makes reliable eviction infeasible, e.g., we observe evictions in only 30% of the attempts. We show that even an informed adversary with a hypothetical way to evict reliably requires linearly more traces in the window size. Our evaluation shows that ObfusCache thwarts state-of-the-art attacks with our recommended window sizes with a large security margin.

Cite

@inproceedings{
  title={{Fast and Secure LLC Caches via Spatial Windows}},
  author={Czerny, Roland and Lammer, Simon and Giner, Lukas and Chakraborty, Anirban and Gruss, Daniel},
  booktitle={DIMVA},
  year={2026},
}