MuSIA: Exploiting multi-source information fusion with abnormal activations for out-of-distribution detection

Abstract

In the open world, out-of-distribution (OOD) detection is crucial to ensure the reliability and robustness of deep learning models. Traditional OOD detection methods are often limited to using single-source information coupled with the abnormal activations of OOD data, resulting in poor detection performance for OOD samples. To this end, we propose MuSIA (Multi-Source Information Fusion with Abnormal Activations) to obtain effective information from multiple information sources and capture abnormal activations to improve the performance of OOD detection. To verify the effectiveness of MuSIA, we conducted experiments with six OOD datasets on six pre-trained models (ViT, RepVGG, DeiT, etc.). Experimental results show that compared with the SOTA method, MuSIA reduces FPR95 ($\downarrow$) by an average of 7.78%. Further ablation studies deeply explore the role of each component in MuSIA, especially the synergy of capturing abnormal activation and multi-source information fusion.