Transforming out timing leaks, more or less

Abstract

We experimentally evaluate program transformations for removing timing side-channel vulnerabilities wrt. security and overhead. Our study of four well-known transformations confirms that their performance overhead differs substantially. A novelty of our work is the empirical investigation of channel bandwidths, which clarifies that the transformations also differ wrt. how much security they add to a program. Interestingly, we observe such differences even between transformations that have been proven to establish timing-sensitive noninterference. Beyond clarification, our findings provide guidance for choosing a suitable transformation for removing timing side-channel vulnerabilities. Such guidance is needed because there is a trade-off between security and overhead, which makes choosing a suitable transformation non-trivial.