Quantifying YOLO Model Robustness: Adversarial Attacks, Common Corruptions, and Bayesian Uncertainty Estimation

Real-world deployment of deep learning object detectors demands resilience against adversarial perturbations, environmental noise corruptions, and out-of-distribution variations. While modern single-stage architectures like YOLO achieve state-of-the-art inference latency and mean average precision on clean benchmarks, systematic post-training forensic auditing remains fragmented. In this work, we present a unified forensic robustness framework integrated into the distributed NeuralForgeAI pipeline. The framework simultaneously quantifies: (1) gradient-based adversarial susceptibility via Fast Gradient Sign Method (FGSM), (2) corruption resilience across five severity tiers under Gaussian blur, Gaussian noise, JPEG compression, and weather variations, and (3) epistemic versus aleatoric uncertainty decomposition using Monte Carlo Dropout (MC Dropout). Evaluating across multi-scale models, our empirical results show that under FGSM perturbations with epsilon = 0.1, model performance drops from 0.820 to 0.533 mAP (35.0% attack success rate), reaching 70.0% degradation at epsilon = 0.2. Common sensor corruptions induce up to 50.6% confidence attenuation at maximum severity.