Peeping tom in the neighborhood: keystroke eavesdropping on multi-user systems

Abstract

A multi-user system usually involves a large amount of information shared among its users. The security implications of such information can never be underestimated. In this paper, we present a new attack that allows a malicious user to eavesdrop on other users ’ keystrokes using such information. Our attack takes advantage of the stack information of a process disclosed by its virtual file within procfs, the process file system supported by Linux. We show that on a multi-core system, the ESP of a process when it is making system calls can be effectively sampled by a “shadow ” program that continuously reads the public statistical information of the process. Such a sampling is shown to be reliable even in the presence of multiple users, when the system is under a realistic workload. From the ESP content, a keystroke event can be identified if they trigger system calls. As a result, we can accurately determine inter-keystroke timings and launch a timing attack to infer the characters the victim entered. We developed techniques for automatically analyzing an application’s binary executable to extract the ESP pattern that fingerprints a keystroke event. The occurrences of such a pattern are identified from an ESP trace the shadow program records from the application’s runtime to calculate timings. These timings are further analyzed using a Hidden Markov Model and other public information related to the victim on a multi-user system. Our experimental study demonstrates that our attack greatly facilitates password cracking and also works very well on recognizing English words. 1