Shoplifting Smart Stores Using Adversarial Machine Learning

Smart stores cashier-less technology is partially based on camera-equipped object detection systems. Powerful machine learning algorithms are deployed at the back-end for classification. In this paper, we explore the usage of adversarial machine learning techniques to deceive the smart stores' classifiers. In particular, we experiment with printable adversarial patches and target making an expensive item classified as a cheaper one. By sticking patches to the objects and lifting them, a customer can make her customized discounts and alter the machine learning prediction. We discuss experiments, results, and possible countermeasures.