Aurora: A Hybrid Deep Learning Approach for Deepfake Detection and Digital Content Authentication
TL;DR
The authors present a hybrid system of Convolutional Neural Networks - Long Short-Term Memory (CNN-LSTM) for an AI-based deepfake and digital authenticity verification system which is more robust and performs better in detection, improving on the traditional CNN-based approaches.
Abstract
The fast development of artificial intelligence tools, specifically–– Generative Adversarial Networks (GANs) and diffusion architecture has empowered the production of synthetic media, more realistic than real, called deepfakes. Readily available doctored images and videos are extremely hazardous to cyber and digital confidences, journalism and privacy. Current detection methods usually do not work with high-quality AI-generated content and the more the false positives the lower the trustworthiness of the results. In this place, the authors present a hybrid system of Convolutional Neural Networks - Long Short-Term Memory (CNN-LSTM) for an AI-based deepfake and digital authenticity verification system. The suggested system is based on the fusion of both spatial feature extraction and sequential analysis as the means of identifying the inconsistencies in manipulated media. It is also provided with a authentication-checking module based on hashing to prevent tampering of content and guarantee integrity. The system comprises of a web built-in real-time detection interface and user interaction interface. Simulation done on a dataset of 84,377 images with an accuracy of 95.2% in validation which is more robust and performs better in detection, improving on the traditional CNN-based approaches. The suggested solution will find practical applications in cybersecurity, digital forensics, media verification, as well as content authentication in the social media.