Cognitive Cyber Crimes in the Era of Artificial Intelligence

4Techniques for Analyzing and Debunking Deepfakes

Alpesh Soni¹*, Kamal Borana², Juber Mirza³ and Abhishek Sharma⁴

¹Department of CSE, SVIIT, SVVV, Indore, India

²Department of AI and DS, SVIIT, SVVV, Indore, India

³Department of CSE, Shri Vaishnav Vidyapeeth Vishwavidyalaya, Indore, India

⁴Department of CSE SVIIT, SVVV, Indore, India

Abstract

The proliferation of deepfakes—synthetic media generated using deep learning—poses severe challenges to information authenticity and digital trust. This study presents a comprehensive approach for analyzing and debunking deepfakes using the FaceForensics++ dataset, which comprises over 1000 original and manipulated videos created using four prominent deepfake generation methods (DeepFakes, Face2Face, FaceSwap, NeuralTextures). Our proposed method, Multifusion DeepFake Detection Network (MF-DFDN), integrates spatiotemporal attention mechanisms, frequency-aware loss, and facial region prioritization to enhance detection performance. Evaluated using precision, recall, F1 score, and a novel deepfake impact factor (DIF), our model achieved an F1 score of 96.7%, precision of 97.3%, recall of 95.1%, and DIF of 0.921, significantly outperforming existing benchmarks. These results affirm MF-DFDN’s robustness in real-world detection tasks and offer a scalable solution for combating digital misinformation.

Keywords: Deepfakes, forgery detection, FaceForensics++, spatiotemporal attention, DIF score

4.1 Introduction

Deepfakes [6, 7] leverage generative ...