Malware white-box evasion attack is a serious threat to machine learning-based malware classification models, where an attacker carefully inserts perturbations into a malware executable at a test time to evade a target model. Previous research introduced different white-box evasion attacks, namely padding and slack attacks, to craft malware adversarial samples and evaluated them based on the perturbation size and their evasion rate against a target model. However, there is a lack of insights into how the malware file structure and type affect the adversarial malware sample generation and their respective evasion rate. In this work, we provide a comprehensive empirical analysis by factoring in the malware structure and the type. Our analysis quantifies slack space availability in various sections, exploring how the slack space can influence the robustness of detection techniques. We further assess the relationship between malware type and evasion rate to understand how different types of malware respond to evasion attacks. Additionally, we explore the connection between each malware type and the corresponding slack space availability, analyzing how these structural factors influence the evasion rates during adversarial attacks. In our experiments, adversarial malware samples were generated using two different algorithms: gradient descent and iterative gradient sign method. This detailed analysis enhances our understanding of evasion dynamics of adversarial attacks across malware types and different structural characteristics of binary malware files.

With the increase in machine learning (ML) applications in different domains, incentives for deceiving these models have reached more than ever. As data is the core backbone of ML algorithms, attackers shifted their interest towards polluting the training data itself. Data credibility is at even higher risk with the rise of state-of-art research topics like open design principles, federated learning, and crowd-sourcing. Since the machine learning model depends on different stakeholders for obtaining data, there are no existing reliable automated mechanisms to verify the veracity of data from each source.Malware detection is arduous due to its malicious nature with the addition of metamorphic and polymorphic ability in the evolving samples. ML has proven to solve the zero-day malware detection problem, which is unresolved by traditional signature- based approaches. The poisoning of malware training data can allow the malware files to go undetected by the ML-based malware detectors, helping the attackers to fulfill their malicious goals. A feasibility analysis of the data poisoning threat in the malware detection domain is still lacking. Our work will focus on two major sections: training ML-based malware detectors and poisoning the training data using the label-poisoning approach. We will analyze the robustness of different machine learning models against data poisoning with varying volumes of poisoning data.