• adversarial attacks
• catastrophic forgetting
• continual learning
• task ordering

The aim of our work is to introduce adversarial data scheduling using the idea of the negative total forward transferability to define an adversarial scheduling that tends to maximize forgetting intentionally. We utilized the Heuristic Continual Task Order Selection (HCTOS) method from "Sequence Transferability and Task Order Selection in Continual Learning" in order to define an adversarial schedule based on the negative transferable metric. The proposed algorithm is the opposite of the HCTOS method.

We have divided our experiments into three different stages. The first stage defines the
Adversarial order at the experience level, an experience represents a set of classes to learn at a time, a CL model learns experiences in a continuous/incremental manner, and in our objective to maximize forgetting, we tried to find the experiences to learn in an order that maximizes forgetting on previously learnt experiences. The second level focuses on the class level, where we craft each experience to contain classes that are difficult to learn, hence increasing forgetting of previously learnt classes, and lastly, in the mini-batch level, we proceed by changing the default data loader to put hard mini-batches, calculated based on average gradient, in an order so that the learning of new experience makes the model struggle to learn and causes an increase in forgetting.

We have also experimented with three different transferability metrics, which are used to calculate the reverse of total forward transferability in order to find an adversarial schedule, the first one we have used is LogME which is used to assess the trained model for transfer learning, the second one is the negative cosine similarity, and the last one is negative gradient similarity.