I did not read the paper fully, just some general commentary, hopefully useful. First, fantastic that you're doing academic work so early. I believe most stude

I did not read the paper fully, just some general commentary, hopefully useful. First, fantastic that you're doing academic work so early. I believe most students wait far too long to be exposed to this aspect of academia (and their lives), which is far more about asking good questions, achieving deep understanding, and getting good results, than memorizing some procedure that isn't necessarily useful (but happens often in school). Keep creative and keep working on creating a good toolset (find math tools you find interesting/useful and own them!). Now on to the paper. I'd reiterate the importance of asking good questions almost above results. For example, the result of adversarial sticks and sinks looks good -- but is it asking the right question? If you think realistically, adversarial attacks can occur in a number of ways. One of them is that a human classifies a dataset one way while a machine another. In this case you would also want the human not to be able to tell you data is weird or there's something funky going on -- that is clearly the case with sticks (and sinks to a lesser extent). A human could be easily trained to spot them, and generally tell something weird is going on. Another attack scenario is where you can modify some object, like a picture, but have some restriction on how much you can modify it. For example, you can manipulate only some bits of an image, or only perturb a small part of a real-world object that is under classification (say by putting a sticker on a car and fooling a system into thinking it is a dog, or something). If there is no restriction on your perturbation, this problem would be trivial (just replace the data with intended object data). The justification behind sticks and sinks does not look very well fundamented. So sticks/sinks do not fare too well in either case, despite looking very good in terms of success vs defenses (although there's a chance they could inspire more practical attacks). The commentary on Haussdorf distance is relevant here, but only on the first case (fooling human judgement), and it is of course an imperfect proxy (the true metric is human perception) -- another hint that fundamentals (and applications) are important to keep in mind. Overall the paper seems well written and I specially like the numerous illustrations. Keep the good work and don't forget to always look for the inspiring, beautiful and impactful, and seeking understanding. With a little of this in mind I have no doubt you can achieve very much. Good luck!