Abstract. This paper describes several novel timing attacks against the common table-driven software implementation of the AES cipher. We define a general attack strategy using a simplified model of the cache to predict timing variation due to cache-collisions in the sequence of lookups performed by the encryption. The attacks presented should be applicable to most high-speed software AES implementations and computing platforms, we have implemented them against OpenSSL v. 0.9.8.(a) running on Pentium III, Pentium IV Xeon, and UltraSPARC III+ machines. The most powerful attack has been shown under optimal conditions to reliably recover a full 128-bit AES key with 2 13 timing samples, an improvement of almost four orders of magnitude over the best previously published attacks of this type [Ber05]. While the task of defending AES against all timing attacks is challenging, a small patch can significantly reduce the vulnerability to these specific attacks with no performance penalty.

This paper describes an algorithm to attack AES using sidechannel information from the final round cache lookups performed by the encryption, specifically whether each access hits or misses in the cache, building off of previous work by Acicmez and Koc [AK06]. It is assumed that an attacker could gain such a trace through power consumption analysis or electromagnetic analysis. This information has already been shown to lead to an effective attack. This paper interprets cache trace data available as binary constraints on pairs of key bytes then reduces key search to a constraint-satisfaction problem. In this way, an attacker is guaranteed to perform as little search as is possible given a set of cache traces, leading to a natural tradeo# between online collection and offline processing. This paper also differs from previous work in assuming a partially pre-loaded cache, proving that cache trace attacks are still effective in this scenario with the number of samples required being inversely related to the percentage of cache which is pre-loaded.