ROSS: A Robust Out-of-Distribution Detection Framework via Synergistic Smoothing
Researchers have introduced ROSS, a novel post-hoc framework designed to enhance the robustness of out-of-distribution (OOD) detection in machine learning systems. Current state-of-the-art OOD detectors are vulnerable to adversarial attacks, compromising their reliability in automated environments. To address this, the authors apply median smoothing to baseline detection scores, effectively balancing clean and adversarial accuracies. The core innovation lies in repurposing noisy samples generated during smoothing to quantify local score instability. The study observes that OOD samples exhibit significantly higher instability under perturbation compared to in-distribution data. Leveraging this insight, ROSS distinguishes between in-distribution and OOD samples with greater precision. Unlike previous methods, ROSS achieves symmetric robustness, defending effectively against both score-minimizing and score-maximizing attacks. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet datasets demonstrate that ROSS outperforms prior methods by up to 40 AUROC points, establishing a new state-of-the-art for robust OOD detection. The accompanying code has been made publicly available to facilitate further research and implementation.
Wire timeline
ROSS: A Robust Out-of-Distribution Detection Framework via Synergistic Smoothing
Researchers have introduced ROSS, a novel post-hoc framework designed to enhance the robustness of out-of-distribution (OOD) detection in machine learning systems. Current state-of-the-art OOD detectors are vulnerable to adversarial attacks, compromising their reliability in automated environments. To address this, the authors apply median smoothing to baseline detection scores, effectively balancing clean and adversarial accuracies. The core innovation lies in repurposing noisy samples generated during smoothing to quantify local score instability. The study observes that OOD samples exhibit significantly higher instability under perturbation compared to in-distribution data. Leveraging this insight, ROSS distinguishes between in-distribution and OOD samples with greater precision. Unlike previous methods, ROSS achieves symmetric robustness, defending effectively against both score-minimizing and score-maximizing attacks. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet datasets demonstrate that ROSS outperforms prior methods by up to 40 AUROC points, establishing a new state-of-the-art for robust OOD detection. The accompanying code has been made publicly available to facilitate further research and implementation.
cs.AI updates on arXiv.org