Malware Analysis and Detection with Explainable Machine Learning

Malware detection is one of the areas where machine learning is successfully employed due to its high discriminating power and the capability of identifying novel variants of malware samples. Typically, the problem formulation is strictly correlated to the use of a wide variety of features covering several characteristics of the entities to classify. Apparently, this practice allows achieving considerable detection performance. However, it hardly permits us to gain insights into the knowledge extracted by the learning algorithm, causing two main issues. First, detectors might learn spurious patterns; thus, undermining their effectiveness in real environments. Second, they might be particularly vulnerable to adversarial attacks; thus, weakening their security. These concerns give rise to the necessity to develop systems that are tailored to the specific peculiarities of the attacks to detect. Within malware detection, Android ransomware represents a challenging yet illustrative domain for assessing the relevance of this issue. Ransomware represents a serious threat that acts by locking the compromised device or encrypting its data, then forcing the device owner to pay a ransom in order to restore the device functionality. Attackers typically develop such dangerous apps so that normally-legitimate components and functionalities perform malicious behaviour; thus, making them harder to be distinguished from genuine applications. In this sense, adopting a well-defined variety of features and relying on some kind of explanations about the logic behind such detectors could improve their design process since it could reveal truly characterising features; hence, guiding the human expert towards the understanding of the most relevant attack patterns. Given this context, the goal of the thesis is to explore strategies that may improve the design process of malware detectors. In particular, the thesis proposes to evaluate and integrate approaches based on rising research on Explainable Machine Learning. To this end, the work follows two pathways. The first and main one focuses on identifying the main traits that result to be characterising and effective for Android ransomware detection. Then, explainability techniques are used to propose methods to assess the validity of the considered features. The second pathway broadens the view by exploring the relationship between explainable machine learning and adversarial attacks. In this regard, the contribution consists of pointing out metrics extracted from explainability techniques that can reveal models' robustness to adversarial attacks, together with an assessment of the practical feasibility for attackers to alter the features that affect models' output the most. Ultimately, this work highlights the necessity to adopt a design process that is aware of the weaknesses and attacks against machine learning-based detectors, and proposes explainability techniques as one of the tools to counteract them.