A typical assumption for deploying machine learning models is that the model training and inference data were drawn from the same distribution. However, this assumption rarely holds true for systems deployed in the open world. Inference data can drift over time for numerous reasons, such as changes in operating conditions, adversarial modifications to targets, or sensor degradation. Despite these changes, deep learning models are especially vulnerable to issuing over-confident predictions on out-of-distribution data. This work seeks to address this issue by proposing a framework for describing out-of-distribution detection pipelines, proposing an out-of-distribution detection algorithm using Gaussian Mixture Models which is well suited for SAR ATR, and by evaluating multiple pipelines which exploit the intermediate states of ATR model deep neural networks. This work studies candidate pipelines with varied amounts of dimensionality reduction and detection algorithms on the SAMPLE+ dataset challenge problems for clutter and confuser rejection. Despite the exclusion of out-of-distribution samples from pipeline training, the presented results demonstrate that these samples can nonetheless be reliably detected, exceeding baseline performance by more than 10 percentage points.
|