Robustness-Aware Fire & Smoke Detection: An XAI-Guided Diagnosis-to-Mitigation Framework

This paper propose an XAI-guided Diagnosis-to-Mitigation (D2M) framework for fire/smoke detection: corruption-specific failures are first diagnosed via Grad-CAM, targeted augmentation is applied only where evidence supports it, and the resulting gains are re-verified through an unbiased XAI attention check rather than accuracy alone. This reveals an asymmetry invisible to standard metrics — Gaussian Noise mitigation is causally verified (+13.78 pp ground-truth-relevant attention), while a similar mAP gain under Low-Light shows no corresponding attention improvement (−1.10 pp) — demonstrating that accuracy alone cannot confirm genuinely improved model reasoning.