Feature-alteration Robustness for Out-of-distribution Detection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Reliable deployment of AI models like neural networks necessitates robust detection and rejection of out-of-distribution (OOD) data.
• Existing methods face challenges in distinguishing between in-distribution (ID) data with altered features and genuine OOD data.

Purpose of the Study:

• To introduce a novel method, Feature-alteration Robustness (FAR), for efficient and reliable detection of OOD samples.
• To demonstrate the effectiveness of measuring feature map robustness under alterations for OOD detection.

Main Methods:

• FAR measures the robustness of intermediate layer feature maps when subjected to specific alterations.
• The method evaluates foreground-background deviations in feature maps across several layers post-alteration.
• Theoretical analysis is provided to elucidate the underlying principles of the FAR method.

Main Results:

• Altered in-distribution (ID) features show robustness in pre-trained models, while OOD features exhibit significant distortion.
• Feature-space distortions in OOD data provide distinctive clues for detection.
• The proposed FAR and FAR+ASH methods achieve state-of-the-art performance on multiple benchmark datasets.

Conclusions:

• Feature map robustness under alterations is a viable indicator for OOD detection.
• FAR offers an efficient and effective approach to enhance AI model reliability by identifying OOD samples.
• The findings suggest a promising direction for developing more dependable AI systems in real-world applications.