Protodetect: An Enhanced Prototype Network for Few-Shot Out-of-Distribution Detection

Out-of-distribution (OOD) detection is essential for identifying data samples that deviate from those encountered during model training, safeguarding the reliability and security of machine learning systems. However, existing few-shot OOD detection approaches face significant limitations. The first category introduces pseudo-unknown samples into training but is heavily dependent on the quality of these samples. The second category constructs distributions solely from known samples, ignoring the potential distribution of unknown samples, which can lead to unclear decision boundaries. This paper introduces Protodetect, a novel metric-learning-based framework designed to overcome these challenges. Leveraging a multi-task learning approach, Protodetect rapidly generalizes learned knowledge to new few-shot OOD scenarios without relying on pseudo-unknown samples. The framework constructs a robust metric space using an improved prototype network, effectively distinguishing in-distribution (ID) samples from OOD samples through precise decision boundaries. We further propose a joint optimization loss function to enhance clustering among same-class samples, improving the clarity of OOD boundaries. Additionally, temperature scaling techniques improve sample separability, thus boosting model accuracy and robustness. Experimental results across multiple datasets confirm that Protodetect achieves superior performance compared to existing few-shot OOD detection methods.