Enhancing Generalization of Universal Adversarial Perturbation through Gradient Aggregation
Deep neural networks are vulnerable to universal adversarial perturbation (UAP), an instance-agnostic perturbation capable of fooling the target model for most samples.