Robust Detection Of Deepfake Videos Using Deep Learning

Abstract

Recent improvements in computational capabilities have significantly propelled the development of deep learning models, making it easier than ever to create synthetic videos that closely resemble real human speech and facial expressions—widely referred to as deepfakes. These convincingly fabricated videos can be exploited for malicious purposes, including political manipulation, staged acts of terrorism, revenge-based pornography, and various forms of digital extortion. To address these threats, this work proposes a deep learning-based system capable of distinguishing real videos from AI-generated deepfakes. The proposed method leverages artificial intelligence to fight against AI-driven deception. Frame-level features are initially extracted from the input video and analyzed using a ResNeXt Convolutional Neural Network. These spatial features are then forwarded to a Recurrent Neural Network architecture powered by Long Short-Term Memory (LSTM) units, which enables the detection of temporal distortions commonly found in manipulated videos. Although the model currently does not support the identification of specific deepfake subtypes such as reenactment or face replacement, it establishes a robust baseline for further innovation in automated deepfake detection.