Actively Enlarging Feature Norms with Universal Adversarial Training and Channel Calibration for Superior OOD Detection

Atik Garg, Yu-Shuen Wang

Keywords: Out-of-distribution (OOD) detection, Adversarial Training

TL;DR: We propose a universal adversarial training framework and a channel-amplified scoring strategy to actively enlarge the feature norm gap between in-distribution and out-of-distribution (OOD) data, significantly improving OOD detection performance.

Abstract: Out-of-distribution (OOD) detection is an increasingly essential component for ensuring the safety and reliability of machine learning systems. A key insight in this area is that the feature norm gap between in-distribution (ID) and OOD data serves as a strong signal for identifying anomalous inputs. Building on this, we propose an innovative Universal Adversarial Training (UAT) framework that actively enlarges this norm gap. Our method introduces a single, learnable Universal Adversarial Map (UAM) that acts as a global regularizer to address shared vulnerable directions in the model. By regularizing the decision boundaries, this approach enlarges the gaps between ID classes, which enhances the model's generalization and its ability to distinguish OOD data. To further enhance detection at inference, we introduce a Channel-Calibrated Feature Norm (CCFN) scoring mechanism that refines the feature norm by suppressing irrelevant background activations. Our comprehensive experiments and ablation studies demonstrate that these innovations lead to substantial performance gains across various benchmarks.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6514