MeDiFakeD: Medical Deepfake Detection using Convolutional Reservoir Networks

Abstract

Generation of photo-realistic fake content using Artificial Intelligence (AI)-based Generative Adversarial Networks has not only engulfed media, facial recognition or social networks, but is now rapidly surging ahead in the realm of medical imaging and is further facilitated by worldwide Covid-19 outbreak. Medical Deepfake pertains to application of AI-triggered deepfake technology on to medical modalities like Computed Tomography (CT) scan, X-Ray, Ultrasound etc. Owing to its high degree of privacy and sensitivity, any threats originating from exposed vulnerabilities, or, attacks on patients medical imagery takes an extremely threatening stance, either devastating the patients remaining lifespan, or resulting in grave financial frauds while satiating corrupt business motives. These tampering attacks, involve either insertion or removal of certain disease conditions, tumors in/from the modality under analysis. This paper implements and demonstrates a practical, lightweight technique which aims to accelerate deepfake detection for biomedical imagery by detecting malignant tumors injected in modalities of healthy patients. The developed technique makes use of convolutional reservoir networks (CoRN), which enable ensemble feature extraction and results in improved classification metrics. We further corroborate its effectiveness while working with a miniscule (< 100) set of images and illustrate the extent of generalization attained with different forms of the same medical imagery.