BOOD: Boundary-based Out-Of-Distribution Data Generation

TL;DR

This work tackles OOD detection by explicitly generating image-level OOD data located around the decision boundaries. It builds a text-conditioned latent space aligned with a diffusion model, then identifies and perturbates ID features closest to the boundary to synthesize informative OOD features that are decoded into images. The approach, BOOD, demonstrates state-of-the-art gains on CIFAR-100 and ImageNet-100, notably reducing FPR95 and increasing AUROC, while maintaining training efficiency comparable to existing diffusion-based methods. By focusing on boundary-adjacent data, BOOD enhances the clarity of ID–OOD boundaries and offers a practical, scalable pathway for improving OOD detection in real-world open-world settings.

Abstract

Harnessing the power of diffusion models to synthesize auxiliary training data based on latent space features has proven effective in enhancing out-of-distribution (OOD) detection performance. However, extracting effective features outside the in-distribution (ID) boundary in latent space remains challenging due to the difficulty of identifying decision boundaries between classes. This paper proposes a novel framework called Boundary-based Out-Of-Distribution data generation (BOOD), which synthesizes high-quality OOD features and generates human-compatible outlier images using diffusion models. BOOD first learns a text-conditioned latent feature space from the ID dataset, selects ID features closest to the decision boundary, and perturbs them to cross the decision boundary to form OOD features. These synthetic OOD features are then decoded into images in pixel space by a diffusion model. Compared to previous works, BOOD provides a more training efficient strategy for synthesizing informative OOD features, facilitating clearer distinctions between ID and OOD data. Extensive experimental results on common benchmarks demonstrate that BOOD surpasses the state-of-the-art method significantly, achieving a 29.64% decrease in average FPR95 (40.31% vs. 10.67%) and a 7.27% improvement in average AUROC (90.15% vs. 97.42%) on the CIFAR-100 dataset.

Figures (8)

• Figure 1: Top: images generated from ID features. Bottom: images generated from OOD features. Compared to preparing ID image datasets, preparing OOD image datasets incurs substantial costs in terms of resource allocation, particularly with respect to labor and time investment. Moreover, certain OOD images, as illustrated in the above figure, are impossible to acquire through real-world data collection methods. Consequently, there exists a pressing need for the development of automated pipelines capable of generating OOD datasets.
• Figure 2: Illustration of perturbing ID boundary feature process. The bar charts under each image represent the prediction probability of the perturbed features by the image classifier. After each perturbation, the prediction probability of the original class decreases. When the prediction of the image classifier switches, we consider the obtained feature crossed the decision boundary.
• Figure 3: Left: the effect of perturbing steps $c$ after crossing the boundary, Right: the effect of step size $\alpha$.
• Figure 4: Illustration of the identified ID boundary features and perturbing them to cross the decision boundary.
• Figure 5: Left: The effect of step size $\alpha$. Right: The effect of perturbation steps $c$ after crossing the boundary.