Quantifying YOLO Model Robustness: Adversarial Attacks, Noise Resilience, and Uncertainty Estimation in Real-World Deployment

YOLO models achieve high mAP under clean laboratory conditions, yet real deployments face sensor noise, weather, and adversarial manipulation. We built an automated post-training robustness pipeline that evaluates YOLOv8n along three axes: (1) FGSM adversarial attack resilience across five perturbation levels, (2) performance degradation under four corruption types at five severity levels, and (3) epistemic and aleatoric uncertainty decomposition via Monte Carlo Dropout with T=20 forward passes. All experiments run on 100 images from COCO128 on a single NVIDIA RTX 4090. Our results show that attack success rates grow from 3.5% (epsilon=0.01) to 70% (epsilon=0.20), that Gaussian noise causes the steepest degradation (up to 46.2% confidence drop at severity 5), and that epistemic variance spans a 24x range across images---demonstrating that MC Dropout can flag low-confidence detections for human review. The pipeline integrates into CI/CD robustness gates and is fully reproducible.