Selective Forgetting in Task-Progressive Learning Through Machine Unlearning

Task-progressive learning (TPL) entails training a model sequentially on multiple tasks while mitigating the phenomenon of 'catastrophic forgetting', where new tasks erase previously learned knowledge. ATPL model can face security challenges, as some tasks are affected by so-called backdoors and adversarial attacks. These attacks exploit vulnerabilities in machine learning models by implanting malicious triggers or perturbing inputs to induce misclassification. This paper explores the role of machine unlearning in TPL to counter those attacks. We consider two common architectures: static and dynamic networks. In static architectures, each task is learned by applying weight update penalties through regularization without any change in neural architecture. In contrast, dynamic architectures expand the network for each task by freezing parameters. We explore the potential for machine unlearning in each scenario to counteract the specified attacks. Specifically, we demonstrate how mechanisms for selective forgetting can be adapted to the TPL models to efficiently 'unlearn' tasks compromised by backdoor and adversarial attacks while preserving the knowledge of other tasks. We demonstrate our method using the MNIST image dataset.