Using Adversarial Training in Developing Robust Models for COVID-19 X-Ray Detection

COVID-19 is a respiratory system disease that has led to a global pandemic and researchers have studied this disease and come up with various solutions. As extra support for Reverse Transcription Polymerase Chain Reaction (RT-PCR) testing or rapid kit testing and to achieve the lowest possible diagnostic error, one of these computerized solutions is using deep learning models for COVID-19 diagnoses through lung X-Ray images. Nevertheless, using advanced technologies introduced advanced risks that require countermeasures such as adversarial attacks The criminal could add some perturbations to the images that cannot be noticed by the human eye but lead the model to misclassify the image. No research is done on developing more secure models for COVID-19 detection, thus in this paper, a new approach is introduced by using neural structured learning to apply adversarial training and train models on perturbed images to make them more robust and secure. After applying this approach to various transfer learning models and experimenting with different values for the hyperparameters, our results have shown that the originally trained models suffer from drawbacks in the performance of 55.25% accuracy on average when tested on perturbed data. On the other hand, the new models that are trained using transfer learning show more robust results with just a drawback of 0.91% in accuracy on average. Such robust models can benefit those who are associated with the medical field.