Robust behavioral malware detection




Kazdagli, Mikhail

Journal Title

Journal ISSN

Volume Title



Computer security attacks evolve to evade deployed defenses. Recent attacks have ranged from exploiting generic software vulnerabilities in memory-unsafe languages such as buffer overflows and format string vulnerabilities to exploiting logic errors in web applications, through means such as SQL injection and cross-site scripting. Furthermore, recent attacks have focused on escalating privileges and stealing sensitive information by exploiting new hardware or operating system (OS) interfaces. Computer security attacks are also now relying on social engineering techniques to run malicious programs on victims' machines; instances of such abuse include phishing and watering hole attacks, both of which trick people into running malicious code or divulging confidential information. Thus, traditional computer security methods, such as OS confinement and program analysis, will not prevent new attacks that do not violate OS confinement or present illegal program behaviors. Another challenge is that traditional security approaches have large trusted code bases (TCBs), which include hardware, OSs, and other software components that implement authentication and authorization logic across a distributed system. This is a vulnerable area because these components are complex and often contain vulnerabilities that undermine the overall system's integrity or confidentiality. Evasive attacks on vulnerable systems -- especially in instances where trusted components turn malicious -- inspire the creation of defenses that can augment formally specified mechanisms against known threats. Specifically, this thesis advances the state of the art in behavioral malware detection -- detecting previously unknown malware in the very early stages of infection within an enterprise network. Here we assess three fundamental insights of modern-day attacks and then describe a cross-layer defense against such attacks. First, we make a low-level machine state visible to behavioral analysis, significantly minimizing the TCB and its associated vulnerabilities. Specifically, our behavioral detector utilizes an executable code's dynamic properties, with architectural and micro-architectural states as input. Second, we evaluate behavioral detectors against adaptive adversaries. For this purpose, we introduce a new metric to determine a detector's robustness against malware modifications, which serves as a step toward explainability of machine learning-based malware detectors. Finally, we exploit the fact that attacks spread through only a limited number of vectors and propose new techniques to analyze the resulting dynamic correlations created among machines. These insights show that behavioral detectors can efficiently protect both individual devices and end hosts within enterprise networks. We present three types of such behavioral detectors. Sherlock protects resource-constrained devices, such as mobile phones and Internet-of-things (IoT) devices, without modifying the software/hardware stack. Sherlock's supervised and unsupervised versions outperform prior work by 24.7% and 12.5% (area under the curve (AUC) metric), respectively, and detects stealthy malware that often evades static analysis tools. The second behavioral detector, Shape-GD, protects devices within an enterprise network. It monitors devices on the network, aggregates data from weak local detectors, overlays that with network-level information, and then makes early, robust predictions regarding malicious activity. Shape-GD achieves its goals by exploiting latent attack semantics. Specifically, it analyzes communication patterns across multiple devices, partitioning them into neighborhoods. Devices within the same neighborhood are likely to be exposed to the same attack vector. Furthermore, we hypothesize that the conditional distribution of false positives is different from that of true positives; i.e., given a neighborhood of nodes, we can compute the aggregate distributional shape of alert feature vectors from the neighborhood itself and provide robust labels. We evaluate Shape-GD by emulating a large community of Windows systems using the system call traces from a few thousand malicious and benign applications; we simulate both a phishing attack in a corporate email network as well as a watering hole attack through a popular website. In both scenarios, Shape-GD identifies malware early on (~100 infected nodes in a ~100K-node system for watering hole attacks, and ~10 of ~1,000 for phishing attacks) and robustly (with ~100% global true-positive and ~1% global false-positive rates). The third behavioral detector, Centurion, detects malware across machines monitored by an anti-virus company. It is able to analyze behavior from 5 million Symantec client machines in real time and discovers malware by correlating file downloads across multiple machines. Compared with a recent local detector that analyzes metadata from file downloads, Centurion reduced the number of false positives from ~1M to ~110K and increased the true-positive rate by a factor of ~2.5. In addition, on average, Centurion detects malware 345 days earlier than commercial anti-virus products.


LCSH Subject Headings