Hacker News Reader

Comment by kough | Hacker News Reader

Super interesting. I'm on mobile and haven't had time to read the whole paper yet - would it be feasible to continuously compute these perturbation vectors during training and include them as part of a larger heuristic? For instance, to incorporate the objective of maximizing the size of the perturbation vector necessary for misclassification? The goal being to end up with a net that is more resistant to such perturbations.

View on HN

Short answer: No. Computing these perturbations requires an expensive optimization with multiple passes through the dataset, and this would be prohibitively expensive to do in the inner-most loop of training.

There are other work in the literature describing faster algorithms to compute these perturbations, which makes it possible to use them while training. See, eg.: https://arxiv.org/abs/1412.6572

IMO, (at least) two pieces of research on the subject means that the short answer really is "yes". Maybe not the exact technique used in the paper in the original post, but conceptually similar techniques.

It's easier to find fooling perturbations of one image, but not of the whole dataset. I assumed the question was can we use the universal perturbations for robust training? The answer to that is still "no", I think.

I suspect that doing this will only get you one step ahead in a cat-and-mouse game. You're only fighting one kind of perturbation.

Reply on news.ycombinator.com