When the Magic Wears Off - Flaws in ML for Security Evaluations and What to Do About It
Explore the flaws in machine learning for security evaluations and learn effective solutions in this conference talk. Delve into the endemic issue of inflated results caused by spatial and temporal biases in academic research on machine learning-based malware classification. Discover a set of space and time constraints for experiment design and a new metric that summarizes classifier performance over time. Examine the TESSERACT open-source evaluation framework, which enables fair comparison of malware classifiers in realistic settings. Gain insights into the distortion of results due to experimental bias and learn about significant improvements achieved through tuning. Cover topics such as cross-validation, temporally inconsistent datasets, time decay, bias from imbalanced testing, and evaluation constraints.