Out-of-Distribution Detection Through Transformation Rectified Activation

Abstract

Rectified activation is a highly effective post-hoc method for out-of-distribution (OOD) detection, designed to enhance the robustness of Deep Neural Networks (DNNs). It operates by truncating high neuron activations to prevent overconfident predictions on unfamiliar data. However, this approach has two key limitations: it only truncates outlier activations without optimizing the post-truncation relationship between in-distribution (ID) and OOD data, and its performance is often unstable due to the manual selection of the rectification threshold. In this paper, we introduce Transformation Rectified Activation (Trans-ReAct), a novel method that addresses these issues. Trans-ReAct leverages two core innovations: 1) it applies mathematical transformations to strategically amplify the activation differentiation between ID and OOD data, and 2) it introduces a dynamic threshold anchor that automatically adapts the rectification threshold to the ID data, ensuring stable performance. Evaluated on three benchmark datasets, Trans-ReAct significantly outperforms state-of-the-art methods, reducing the false positive rate at 95% true positive rate (FPR95) by up to 9.76% and increasing the Area Under the ROC Curve (AUROC) by up to 2.04%.