The current studies of provable robustness for deep neural networks (DNNs) usually assume that the class distribution is overall balanced. However, in real-world applications especially for safety-sensitive systems, the class distribution often exhibits a long... ...
