Model collapse occurs when there is a trivial solution to an optimization problem and the model is trained to use that trivial solution. For example, if you want to minimize distance between learned representations of data augmentations for the same image (like for ), the easiest way is to map everything to a constant vector. The adversarial attacker could always map an input to random noise to optimally mess with the predictor model, but that would not help the predictor model learn and build robustness. There are varieties of methods to prevent model collapse, and they are all very model- and problem-specific.