Firenze: Model Evaluation Using Weak Signals

Data labels in the security field are frequently noisy, limited, or biased towards a subset of the population. As a result, commonplace evaluation methods such as accuracy, precision and recall metrics, or analysis of performance curves computed from labeled datasets do not provide sufficient confidence in the real-world performance of a machine learning (ML) model. This has slowed the adoption of machine learning in the field. In the industry today, we rely on domain expertise and lengthy manual evaluation to build this confidence before shipping a new model for security applications. In this paper, we introduce Firenze, a novel framework for comparative evaluation of ML models' performance using domain expertise, encoded into scalable functions called markers. We show that markers computed and combined over select subsets of samples called regions of interest can provide a robust estimate of their real-world performances. Critically, we use statistical hypothesis testing to ensure that observed differences-and therefore conclusions emerging from our framework-are more prominent than that observable from the noise alone. Using simulations and two real-world datasets for malware and domain-name-service reputation detection, we illustrate our approach's effectiveness, limitations, and insights. Taken together, we propose Firenze as a resource for fast, interpretable, and collaborative model development and evaluation by mixed teams of researchers, domain experts, and business owners.

Adversarial Robustness with Non-uniform Perturbations

Robustness of machine learning models is critical for security related applications, where real-world adversaries are uniquely focused on evading neural network based detectors. Prior work mainly focus on crafting adversarial examples with small uniform norm-bounded perturbations across features to maintain the requirement of imperceptibility. Although such approaches are valid for images, uniform perturbations do not result in realistic adversarial examples in domains such as malware, finance, and social networks. For these types of applications, features typically have some semantically meaningful dependencies. The key idea of our proposed approach is to enable non-uniform perturbations that can adequately represent these feature dependencies during adversarial training. We propose using characteristics of the empirical data distribution, both on correlations between the features and the importance of the features themselves. Using experimental datasets for malware classification, credit risk prediction, and spam detection, we show that our approach is more robust to real-world attacks. Our approach can be adapted to other domains where non-uniform perturbations more accurately represent realistic adversarial examples.