Deep Neural Networks (DNNs) have demonstrated exceptional accuracy in a wide range of computer vision applications such as autonomous driving and healthcare. Different kinds of DNNs, including Convolution Neural Networks (CNNs), Vision transformers, and hybrid CNN-Transformers, have been utilized for these applications. Nevertheless, as these models grow in complexity and demand more computational resources, deploying them effectively across diverse edge devices with varying efficiency constraints becomes increasingly challenging. For instance, the deployment of a transformer network on a mobile device with constrained memory capacity is challenging, just as the deployment of CNNs on a microcontroller with a mere 256KB of memory. Therefore, it is necessary to manually or automatically design a DNN for each edge device which is a very costly and environmental endeavor (CO2 emission) task. Moreover, these applications are safety-critical, so the security aspect of DNNs is crucial for them. For example, adversarial attacks characterized by subtle input perturbations are able to fool DNNs deployed in these applications and cause them to make incorrect decisions. In order to have a robust model, we need to use time-consuming adversarial training methods that add more complexity to the training mechanism of the DNNs and compound the cost and CO2 emission challenges. Although designing and training a robust DNN for each hardware platform is prohibitively expensive, it is required for safety-critical applications. The primary objective of this proposal is to design a robust neural architecture search framework to create a supernetwork based on CNNs, transformers, and hybrid CNN-transformers that can be robust against adversarial attacks and support a variety of deployment scenarios.