Evaluating Explainability in Safety-Critical ATR Systems
A new arXiv paper (2605.05748) evaluates explainability methods for safety-critical Automatic Target Recognition (ATR) systems. The authors identify major XAI paradigms—saliency-based, attention-based, surrogate approaches, and detection-aware extensions—and formalize explainability as an assurance-oriented assessment. They introduce a taxonomy and assess methods across four dimensions: interpretability, robustness, vulnerability to manipulation, and suitability for validation and verification. The paper argues that high predictive performance alone is insufficient for ATR systems operating on image, video, radar, and multisensor data.
Key facts
- arXiv paper 2605.05748 evaluates explainability in safety-critical ATR systems.
- Identifies XAI paradigms: saliency-based, attention-based, surrogate, and detection-aware.
- Formalizes explainability as an assurance-oriented assessment problem.
- Introduces a taxonomy for XAI methods in ATR.
- Assesses methods on interpretability, robustness, manipulation vulnerability, and validation suitability.
- ATR systems use image, video, radar, and multisensor data.
- High predictive performance alone is deemed insufficient for safety-critical ATR.
- The paper is published on arXiv.
Entities
Institutions
- arXiv