SNIFFER: A high-accuracy malware detector for enterprise-based systems

Abstract

In the continual battle between malware attacks and antivirus technologies, both sides strive to deploy their techniques at always lower layers in the software system stack. The goal is to monitor and control the software executing in the levels above their own deployment, to detect attacks or to defeat defenses. Recent antivirus solutions have gone even below the software, by enlisting hardware support. However, so far, they have only mimicked classic software techniques by monitoring software clues of an attack. As a result, malware can easily defeat them by employing metamorphic manifestation patterns. With this work, we propose a hardware-monitoring solution, SNIFFER, which tracks malware manifestations in system-level behavior, rather than code patterns, and it thus cannot be circumvented unless malware renounces its very nature, that is, to attack. SNIFFER leverages in-hardware feature monitoring, and uses machine learning to assess whether a system shows signs of an attack. Experiments with a virtual SNIFFER implementation, which supports 13 features and tests against five common network-based malicious behaviors, show that SNIFFER detects malware nearly 100% of the time, unless the malware aggressively throttle its attack. Our experiments also highlight the need for machine-learning classifiers employing a range of diverse system features, as many of the tested malware require multiple, seemingly disconnected, features for accurate detection.