Are Machine Learning Classifiers Robust Enough Against Poisoned Datasets?

Machine learning (ML) classifiers are widely used in many real-world applications for classification, prediction, and regression tasks. However, data characteristics (e.g., class imbalance, feature overlap, and superfluous augmentation) and inadequate logic inside ML classifiers for catering to data-related problems often lead to undesirable outcomes. This article delves into ML failure when data are poisoned through adversarial perturbations, which can be intentional and/or malicious. Specifically, we perturb the data composition/structure and analyze the robustness of ML classifiers in dealing with them. We found that most ML classifiers cannot withstand data perturbations, and they produce benign-looking but undesirable results. We underscore the causes of such failures, their implications for the ML field, and how to prevent them. Our work contributes to making next-generation ML classifiers more robust and reliable in the upcoming artificial intelligence (AI)-driven era.