Mittal, PrateekJacob, Dennis2024-07-032026-09-292024-07-032026-09-292024-04-152024-07-03http://arks.princeton.edu/ark:/88435/dsp013197xq389https://theses-dissertations.princeton.edu/handle/88435/dsp013197xq389The advent of deep learning has brought about vast improvements to computer vision systems and enabled technologies such as self-driving cars, facial recognition, etc. Nevertheless, these models have been found to be susceptible to adversarial attacks. Of particular importance to the research community are patch attacks, which have been found to be realizable in the physical world. As a result, researchers have proposed a variety of defense mechanisms in order to circumvent patch attacks. A security "arms race" between attackers and defenders has made certifiable defenses for patch attacks, which feature provable guarantees on robustness, especially valuable in the ML security community. While certifiable defenses like PatchCleanser and ObjectSeeker have been successful at providing guarantees on robustness in the single-label classification and object detection domains respectively, less work has been done on proposing a certifiable defense for patch attacks in the multi-label classification domain. To this end, we propose an extension to PatchCleanser for the multi-label classification domain called Multi-Label PatchCleanser. By constructing an inference algorithm and certification procedure in tandem, we are able to define notions of robustness for the multi-label classification setting based on precision and recall. We find that Multi-Label PatchCleanser can achieve non-trivial robustness on the MSCOCO 2014 validation dataset while maintaining high clean performance; this can be augmented by techniques such as cutout pre-training and alternative computer vision backbones. Additionally, we discover a promising insight about mask augmentations in the multi-label classification domain which provides improvements to our current baseline robustness.application/pdfenTowards a provably-certifiable defense for multi-label classifiers against adversarial patchesPrinceton University Senior Theses