Imperceptible backdoor triggers induce false memories in continual learning models

Attackers can inject backdoor samples imperceptible to the human eye during continual learning model training, causing the model to forget specific tasks or classes at test time, forming "false memories." The attack targets generative replay and regularization-based continual learning methods, and is validated as effective on MNIST variants, SVHN, and CIFAR 10. The backdoor pattern can be injected at any time point, and success requires only adding 1% of the total training data of a single possibly unrelated task.