Adversary Aware Continual Learning (AACL): a defense framework against adversarial backdoor attacks in class-incremental learning

The researchers propose the Adversary Aware Continual Learning (AACL) defense framework to defend against adversarial backdoor attacks that cause class-incremental learning models to forget specific tasks or classes at test time. The framework learns perceivable stronger patterns during training to suppress imperceptible backdoor patterns injected by the attacker, without needing to know the attacker's target task, target class, or the shape, size and location of the backdoor pattern. The defense effectiveness has been validated on replay-based class-incremental learning algorithms and continual learning benchmark variants of CIFAR-10, CIFAR-100 and MNIST.