Safety-critical autonomous systems utilizing deep learning for object detection are vulnerable to input perturbations that can compromise operational reliability. Distinguishing between naturally occurring sensor disturbances and adversarial attacks is essential for implementing appropriate defensive measures. We propose SAD (Statistically significant Adversarial Detection), an approach that uses Mahalanobis distance analysis of bounding box coordinates and class feature embeddings to identify patterns characteristic of intentional manipulation versus non-directed random perturbations. The approach is illustrated using synthetic data relevant to a chaser-target docking satellite case study employing an SSDLite object detector. We compare white-box Projected Gradient Descent attacks, black-box Zero-Order Optimization attacks, equivalent-magnitude uniform noise, and domain-relevant sunspot anomalies. Our results show that adversarial perturbations produce statistically significant deviations in Mahalanobis distance metrics at perturbation levels where i) model performance begins to degrade while comparable random noise, and sunspot anomalies do not induce such degradation nor trigger SAD detection; and ii) ‘natural’ anomaly detection (using variational autoencoder reconstruction error) does not occur for adversarials but does for sunspot anomalies. The proposed approach enables distinctions to be made at the level of anomaly detection within Operation and Monitoring architectures, supporting informed decision-making for autonomous space systems where identifying attack vectors is critical for mission safety.