Towards Certified Object Detectors: Certified Runway Detection Using Yolo

In recent years, significant progress has been made in certifying the robustness of image classifiers. However, direct certification for object detectors remains largely unexplored. Object detection adds complexities like precise localization and managing multiple loss functions. While object detectors like YOLO perform well on natural samples, even state-of-the-art models are susceptible to minimal adversarial perturbations, limiting their deployment in safety-critical areas. This work presents the first framework for the direct certification of object detection using a YOLO-based architecture. By employing a modified YOLOv2 architecture and interval bound propagation (IBP) training, we achieved certified runway detection on a scaled-down (128 × 128) LARD dataset.