Minimum Selection Feature Importance Guided Attack

Machine learning models are under severe threat by adversarial attacks. Machine learning is swiftly taking on a central role in organizations' value propositions, and as a result, organizations need to safeguard them. Like every machine learning model, adversarial attack algorithms are trained on data. Recently, specific a I gorithms have been developed for tabular datasets such as the Feature Importance Guided Attack (FIGA) [1]. However, FIGA assumes the knowledge of the total available data for training. In a typical case, the attacker will query a target model (an oracle) to label their training data and perform an adversarial attack on the labelled data. Attackers would like to minimize the queries made in order to remain discreet; however, FIGA requires the maximum amount of data. We present a novel methodology to train FIGA with reduced training samples. This method leverages an intuitive approach that utilizes the K-means clustering algorithm to select legitimate samples from the training data for querying. Our results demonstrate a 99 % reduction in the data required for training.