When the Magic Wears Off: Flaws in ML for Security Evaluations (and What to Do about It)

Academic research on machine learning-based malware classification appears to leave very little room for improvement, boasting F1 performance figures of up to 0.99. Is the problem solved? In this talk, we argue that there is an endemic issue of inflated results due to two pervasive sources of experimental bias: spatial bias, caused by distributions of training and testing data not representative of a real-world deployment, and temporal bias, caused by incorrect splits of training and testing sets (e.g., in cross-validation) leading to impossible configurations. To overcome this issue, we propose a set of space and time constraints for experiment design. Furthermore, we introduce a new metric that summarizes the performance of a classifier over time, i.e., its expected robustness in a real-world setting. Finally, we present an algorithm to tune the performance of a given classifier. We have implemented our solutions in TESSERACT, an open source evaluation framework that allows a fair comparison of malware classifiers in a realistic setting. We used TESSERACT to evaluate two well-known malware classifiers from the literature on a dataset of 129K applications, demonstrating the distortion of results due to experimental bias and showcasing significant improvements from tuning.