NGA Project

Using Using Deep Learning to Identify Deepfakes in Safelite Imagery


In recent years, the proliferation of "deepfakes" or realistic looking computer generated imagery has become not only an issue with social implications, but also with national security ones. For this project, Edwin Purcell, Stephen Whetzel and I worked with the NGA to create a deep learning model for detecting deepfake satellite images. On a publically available set of data, we were able to identify which images were authentic and which were fake with over 90% accuracy. We then used saliency mapping to analyze exactly which sections of images were most important for our model to distinguish between real and authentic.