Impact of Deepfakes and Gradient Centralization on COVID-19 Diagnosis: A Compact Swin Transformer-Based Approach

In this paper, a hierarchical vision transformer with sliding window attention (Swin transformer) has been trained with and without synthetic COVID-19-infected frontal Chest X-Ray (CXR) images in identifying the disease. The architecture is trained, validated, and tested using different proportions of a cumulative dataset comprising 30, 386 images. In the implementation, some of the COVID-19-infected real chest X-rays are replaced with synthetic X-ray images for identifying the effect of deepfakes (DF) on COVID-19 diagnosis. It is observed that the inclusion of DFs not only improved the scores and smoothened the learning curves with the help of gradient centralized RMSProp optimizer but also achieved a standardized accuracy of \(93.5\%\) with \(96.50\%\) sensitivity and 0.035 FNR in detecting COVID-19 from CXR images.