Biblio

Cybersecurity plays a critical role in protecting sensitive information and the structural integrity of networked systems. As networked systems continue to expand in numbers as well as in complexity, so does the threat of malicious activity and the necessity for advanced cybersecurity solutions. Furthermore, both the quantity and quality of available data on malicious content as well as the fact that malicious activity continuously evolves makes automated protection systems for this type of environment particularly challenging. Not only is the data quality a concern, but the volume of the data can be quite small for some of the classes. This creates a class imbalance in the data used to train a classifier; however, many classifiers are not well equipped to deal with class imbalance. One such example is detecting malicious HMTL files from static features. Unfortunately, collecting malicious HMTL files is extremely difficult and can be quite noisy from HTML files being mislabeled. This paper evaluates a specific application that is afflicted by these modern cybersecurity challenges: detection of malicious HTML files. Previous work presented a general framework for malicious HTML file classification that we modify in this work to use a $\chi$2 feature selection technique and synthetic minority oversampling technique (SMOTE). We experiment with different classifiers (i.e., AdaBoost, Gentle-Boost, RobustBoost, RusBoost, and Random Forest) and a pure detection model (i.e., Isolation Forest). We benchmark the different classifiers using SMOTE on a real dataset that contains a limited number of malicious files (40) with respect to the normal files (7,263). It was found that the modified framework performed better than the previous framework's results. However, additional evidence was found to imply that algorithms which train on both the normal and malicious samples are likely overtraining to the malicious distribution. We demonstrate the likely overtraining by determining that a subset of the malicious files, while suspicious, did not come from a malicious source.

Machine learning and data mining algorithms typically assume that the training and testing data are sampled from the same fixed probability distribution; however, this violation is often violated in practice. The field of domain adaptation addresses the situation where this assumption of a fixed probability between the two domains is violated; however, the difference between the two domains (training/source and testing/target) may not be known a priori. There has been a recent thrust in addressing the problem of learning in the presence of an adversary, which we formulate as a problem of domain adaption to build a more robust classifier. This is because the overall security of classifiers and their preprocessing stages have been called into question with the recent findings of adversaries in a learning setting. Adversarial training (and testing) data pose a serious threat to scenarios where an attacker has the opportunity to ``poison'' the training or ``evade'' on the testing data set(s) in order to achieve something that is not in the best interest of the classifier. Recent work has begun to show the impact of adversarial data on several classifiers; however, the impact of the adversary on aspects related to preprocessing of data (i.e., dimensionality reduction or feature selection) has widely been ignored in the revamp of adversarial learning research. Furthermore, variable selection, which is a vital component to any data analysis, has been shown to be particularly susceptible under an attacker that has knowledge of the task. In this work, we explore avenues for learning resilient classification models in the adversarial learning setting by considering the effects of adversarial data and how to mitigate its effects through optimization. Our model forms a single convex optimization problem that uses the labeled training data from the source domain and known- weaknesses of the model for an adversarial component. We benchmark the proposed approach on synthetic data and show the trade-off between classification accuracy and skew-insensitive statistics.

Feature selection is an important step in data analysis to address the curse of dimensionality. Such dimensionality reduction techniques are particularly important when if a classification is required and the model scales in polynomial time with the size of the feature (e.g., some applications include genomics, life sciences, cyber-security, etc.). Feature selection is the process of finding the minimum subset of features that allows for the maximum predictive power. Many of the state-of-the-art information-theoretic feature selection approaches use a greedy forward search; however, there are concerns with the search in regards to the efficiency and optimality. A unified framework was recently presented for information-theoretic feature selection that tied together many of the works in over the past twenty years. The work showed that joint mutual information maximization (JMI) is generally the best options; however, the complexity of greedy search for JMI scales quadratically and it is infeasible on high dimensional datasets. In this contribution, we propose a fast approximation of JMI based on information theory. Our approach takes advantage of decomposing the calculations within JMI to speed up a typical greedy search. We benchmarked the proposed approach against JMI on several UCI datasets, and we demonstrate that the proposed approach returns feature sets that are highly consistent with JMI, while decreasing the run time required to perform feature selection.