Reconnaissance Dataset

Secure speculation schemes are critical for defeat-ing speculative side-channel attacks, but often incur notableperformance penalties, encouraging research into optimizationstrategies. One such strategy, ReCon, exploits previous non-speculative data leakage to remove protections for already leakeddata. We show that ReCon leaks potential secrets in cases wherean instruction has multiple secret sources. We present insightsinto the interaction of Speculative Taint Tracking (STT) andReCon, and show how their combination undermines STT\u2019ssecurity guarantee. We present ReConnaissance, which fixes thisoversight, and show that it lowers the performance overhead-reduction of ReCon from 23.5% to 18.3%.