Catastrophic Learning: Poisoned Data Can Block a Continual-Learning Model From Acquiring Knowledge It Has Not Seen Yet

arXiv 2608.18976·medium signalXBlueskyLinkedInCopy link

Adversarial work on continual learning has mostly aimed at re-triggering catastrophic forgetting. arXiv 2608.18976 (2026-08-19) identifies the opposite flaw, attacking plasticity rather than stability: 'learning blockers' are manipulated data that reduce the learnability of current or upcoming iterations, and they are hard to detect during the current iteration because they can target iterations whose data the model has not encountered. The paper formalizes six attack strategies (Label-Exchange, Tensor-Exchange, plus Attraction and Repulsion variants in Coincident and Preceding forms) and evaluates on MNIST and Split-CIFAR10 against DER, ER-ACE and iCaRL across more than 4,480 simulations, showing an adversary can simultaneously impede new learning and promote loss of prior knowledge.