Skip to content

ATOM: Enhancing OOD Detection Robustness with Outlier Mining

ATOM: Robustifying Out-of-distribution Detection Using Outlier Mining

arXiv· arXiv:2006.15207· ECML/PKDD· Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, Somesh JhaDepartment of Computer Sciences, University of Wisconsin-Madison; GoogleIndexed by this site · 23 days ago· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryATOM improves the robustness of OOD detectors against adversarial OOD inputs by mining informative auxiliary OOD samples and combining them with adversarial training. The authors' analysis finds that most auxiliary OOD samples may be unhelpful or even harm the detector's decision boundary. On CIFAR-10, ATOM reduces the FPR at 95% TPR under adversarial OOD inputs by up to 57.99%, achieves SOTA in classic and adversarial OOD evaluations, and generalizes to unseen adversarial attacks.

Questions
Existing OOD detection methods are fragile under adversarial perturbations: the authors report that on CIFAR-10 the FPR of OE rises from 3.66% to 99.94% under attack. The authors consider that randomly sampled auxiliary outliers are mostly uninformative and cannot improve, and may even harm, the detection boundary.
Methods
Proposes ATOM: use a (K+1)-class classifier for adversarial training (PGD solves the inner max), each epoch sample N points from the auxiliary OOD pool, sort by OOD score, and select n informative samples starting from the qN-th for training; also provides a theoretical analysis under a Gaussian data model showing that outlier mining reduces ID sample complexity.
Experiments and results
On CIFAR-10/100 and SVHN, 8 methods are compared across four types of OOD inputs (Natural, Corruption, L∞, Comp.). On CIFAR-10, the FPR of ATOM is: Natural 1.69%, Corruption 25.26%, L∞ 20.55%, Comp. 38.89%, all better than the baselines; on SVHN's Comp. OOD, ATOM is 96.33%, close to failure.
Limitations and future work
The paper has no Limitations section discussing limitations. Experimental coverage: the ID datasets are SVHN/CIFAR-10/CIFAR-100, the architectures are DenseNet and WideResNet, and the auxiliary data are TinyImages and ImageNet-RC; all experiments are run only once (with a fixed random seed), and the paper does not report variance across multiple runs.

Related papers

Show 3 more papers
Read originalECML/PKDD