Towards effective masquerade attack detection

Abstract

Data theft has been the main goal of the cybercrime community for many years, and more and more so as the cybercrime community gets more motivated by financial gain establishing a thriving underground economy. Masquerade attacks are a common security problem that is a consequence of identity theft and that is generally motivated by data theft. Such attacks are characterized by a system user illegitimately posing as another legitimate user. Prevention-focused solutions such as access control solutions and Data Loss Prevention tools have failed in preventing these attacks, making detection not a mere desideratum, but rather a necessity. Detecting masqueraders, however, is very hard. Prior work has focused on user command modeling to identify abnormal behavior indicative of impersonation. These approaches suffered from high miss and false positive rates. None of these approaches could be packaged into an easily-deployable, privacy-preserving, and effective masquerade attack detector. In this thesis, I present a machine learning-based technique using a set of novel features that aim to reveal user intent. I hypothesize that each individual user knows his or her own file system well enough to search in a limited, targeted, and unique fashion in order to find information germane to their current task. Masqueraders, on the other hand, are not likely to know the file system and layout of another user's desktop, and would likely search more extensively and broadly in a manner that is different from that of the victim user being impersonated. Based on this assumption, I model a user's search behavior and monitor deviations from it that could indicate fraudulent behavior. I identify user search events using a taxonomy of Windows applications, DLLs, and user commands. The taxonomy abstracts the user commands and actions and enriches them with contextual information. Experimental results show that modeling search behavior reliably detects all simulated masquerade activity with a very low false positive rate of 1.12%, far better than any previously published results. The limited set of features used for search behavior modeling also results in considerable performance gains over the same modeling techniques that use larger sets of features, both during sensor training and deployment. While an anomaly- or profiling-based detection approach, such as the one used in the user search profiling sensor, has the advantage of detecting unknown attacks and fraudulent masquerade behaviors, it suffers from a relatively high number of false positives and remains potentially vulnerable to mimicry attacks. To further improve the accuracy of the user search profiling approach, I supplement it with a trap-based detection approach. I monitor user actions directed at decoy documents embedded in the user's local file system. The decoy documents, which contain enticing information to the attacker, are known to the legitimate user of the system, and therefore should not be touched by him or her. Access to these decoy files, therefore, should highly suggest the presence of a masquerader. A decoy document access sensor detects any action that requires loading the decoy document into memory such as reading the document, copying it, or zipping it. I conducted human subject studies to investigate the deployment-related properties of decoy documents and to determine how decoys should be strategically deployed in a file system in order to maximize their masquerade detection ability. Our user study results show that effective deployment of decoys allows for the detection of all masquerade activity within ten minutes of its onset at most. I use the decoy access sensor as an oracle for the user search profiling sensor. If abnormal search behavior is detected, I hypothesize that suspicious activity is taking place and validate the hypothesis by checking for accesses to decoy documents. Combining the two sensors and detection techniques reduces the false positive rate to 0.77%, and hardens the sensor against mimicry attacks. The overall sensor has very limited resource requirements (40 KB) and does not introduce any noticeable delay to the user when performing its monitoring actions. Finally, I seek to expand the search behavior profiling technique to detect, not only malicious masqueraders, but any other system users. I propose a diversified and personalized user behavior profiling approach to improve the accuracy of user behavior models. The ultimate goal is to augment existing computer security features such as passwords with user behavior models, as behavior information is not readily available to be stolen and its use could substantially raise the bar for malefactors seeking to perpetrate masquerade attacks.