This reminds me of how, if I recall correctly, in the original paper on adversarial attacks, the authors found that adversarial attacks on one neural network would generally have some success on other neural networks if they were trained for similar tasks (say, labeling images).
No comments yet.