Current semantic segmentation models are vulnerable to the addition of imperceptible perturbations to the input data [1]. These perturbations are well-crafted attacks that when added in small amounts to a sample, drastically fool the model and decrease accuracy [2]. To combat this challenge and make the model robust toward attacks, adversarial training is performed on the model. To perform adversarial training on a model, one generates the adversarial examples and then either continually trains the trained model, or mixes the attack examples with non-attacked examples and trains a new model from scratch [3, 4]. In this approach, the model performs worse on each individual data attack type due to the updated weights, but better overall. Another approach is to detect the attacked data to avoid it during classification via adversarial detection. In this case, the attacks are often ignored, despite having information that might be valuable to classify [5].