Identifying Deepfake Faces with ResNet50-Keras using Amazon EC2 DL1 Instances powered by Gaudi Accelerators from Habana Labs

With the emergence of deepfake technology, it has become exponentially more difficult to identify real content from the artificially generated on social media. To counter the ill-effects of online deepfake content, we propose an application to predict if a given facial image is real or generated virtually via deepfake technology. We train a custom ResNet50-Keras model on the new AWS EC2 DL1 Instances powered by Gaudi accelerators developed by Habana Labs (an Intel company) and integrate the instance with Amazon's S3 bucket. With the model saved as an h5 file, we make a RestAPI using FastAPI. The API takes an input image, converts it into a JSON request, and passes it to the backend. Faces extracted from these images are passed through various pre-processing methods and finally to the model that classifies them to be either a deepfake or not and generates a face mesh accordingly. The model combines the processed faces into a single image sent back as a response to the client that changes the stateHooks to display the desired result. Further, we have summarized the results obtained when detecting such manipulated images.