Detecting Out-of-distribution Examples via Class-conditional Impressions Reappearing
This addresses privacy and security concerns in real-world scenarios by enabling OOD detection without auxiliary data, though it is incremental as it builds on existing post-hoc methods.
The paper tackles out-of-distribution detection without needing training data by proposing C2IR, a data-free method that uses image impressions to recover feature statistics, achieving performance comparable to full-access methods on datasets like SVHN.
Out-of-distribution (OOD) detection aims at enhancing standard deep neural networks to distinguish anomalous inputs from original training data. Previous progress has introduced various approaches where the in-distribution training data and even several OOD examples are prerequisites. However, due to privacy and security, auxiliary data tends to be impractical in a real-world scenario. In this paper, we propose a data-free method without training on natural data, called Class-Conditional Impressions Reappearing (C2IR), which utilizes image impressions from the fixed model to recover class-conditional feature statistics. Based on that, we introduce Integral Probability Metrics to estimate layer-wise class-conditional deviations and obtain layer weights by Measuring Gradient-based Importance (MGI). The experiments verify the effectiveness of our method and indicate that C2IR outperforms other post-hoc methods and reaches comparable performance to the full access (ID and OOD) detection method, especially in the far-OOD dataset (SVHN).