FROB: Few-shot ROBust Model for Classification and Out-of-Distribution Detection

Nowadays, classification and Out-of-Distribution (OoD) detection in the\nfew-shot setting remain challenging aims due to rarity and the limited samples\nin the few-shot setting, and because of adversarial attacks. Accomplishing\nthese aims is important for critical systems in safety, security, and defence.\nIn parallel, OoD detection is challenging since deep neural network classifiers\nset high confidence to OoD samples away from the training data. To address such\nlimitations, we propose the Few-shot ROBust (FROB) model for classification and\nfew-shot OoD detection. We devise FROB for improved robustness and reliable\nconfidence prediction for few-shot OoD detection. We generate the support\nboundary of the normal class distribution and combine it with few-shot Outlier\nExposure (OE). We propose a self-supervised learning few-shot confidence\nboundary methodology based on generative and discriminative models. The\ncontribution of FROB is the combination of the generated boundary in a\nself-supervised learning manner and the imposition of low confidence at this\nlearned boundary. FROB implicitly generates strong adversarial samples on the\nboundary and forces samples from OoD, including our boundary, to be less\nconfident by the classifier. FROB achieves generalization to unseen OoD with\napplicability to unknown, in the wild, test sets that do not correlate to the\ntraining datasets. To improve robustness, FROB redesigns OE to work even for\nzero-shots. By including our boundary, FROB reduces the threshold linked to the\nmodel's few-shot robustness; it maintains the OoD performance approximately\nindependent of the number of few-shots. The few-shot robustness analysis\nevaluation of FROB on different sets and on One-Class Classification (OCC) data\nshows that FROB achieves competitive performance and outperforms benchmarks in\nterms of robustness to the outlier few-shot sample population and variability.\n