WOOD: Wasserstein-Based Out-of-Distribution Detection

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks often exhibit high-confidence predictions on out-of-distribution (OOD) samples, posing risks in critical applications.
• Detecting OOD samples is vital for enhancing classifier robustness, resilience, and security against adversarial attacks and irrelevant inputs.

Purpose of the Study:

• To develop a novel method for detecting OOD samples that is compatible with various classifier architectures.
• To address the challenges of OOD detection, including multiple unknown distributions and the need for effective score functions.

Main Methods:

• Proposes a Wasserstein-based out-of-distribution detection (WOOD) method.
• Defines a Wasserstein-based score to quantify the dissimilarity between test samples and in-distribution data.
• Formulates and solves an optimization problem using the proposed score function.

Main Results:

• The WOOD method demonstrates consistent superior performance compared to existing OOD detection techniques.
• Statistical learning bounds were investigated, ensuring the empirical optimizer approximates the global optimum.

Conclusions:

• The proposed WOOD method offers an effective and compatible solution for OOD sample detection.
• This approach enhances the reliability and security of deep learning models in real-world scenarios.