We present a new Generative Adversarial Network Architecture for perspective augmented image samples which can effectively discriminate over physical attacks and recover contextual information through our dual-stage GAN architecture. Using a novel method known as Adversarial Sample Synthesis, our sampling strategy trains on pairs of images where we utilize an adversarial sub-sampling approach to effectively learn the divergence between the fake and real image distributions. Our methods have shown to effectively discriminating between fakes and real images while also doubling as a generative model which can perform an inferential recovery process to predict the entire context of the image. Our model requires very little hyper-parameter tuning and converges relatively fast, thus yielding a low-cost for training and saving energy. Although our results show great promise, we will not release the underlying model to consider ethics and surrounding consequences that may emerge from the misuse of this model.