Exploring lightweight deep learning techniques for efficient deepfake detection

• Main Author: Chua, Gim Aik (Cai JinYi)
• Other Authors: Deepu Rajan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175361
• Institution: Nanyang Technological University

Deepfakes refer to a form of synthetic media where artificial intelligence and deep learning techniques are used to create or manipulate audio and video content to depict individuals saying or doing things they never actually said or did. Deepfake detection is the ability to differentiate between deepfakes and real media. In this project, we undertake the task of proposing strategies to achieve high deepfake detection accuracy, while attempting to reduce model complexity, size, and inference time, with the goal of pushing the deployment of complex deep-network based solutions such as EfficientNet closer towards hardware constrained environments such as edge devices. In this project, we propose strategies to tackle the deepfake video detection problem and learning strategies that achieved high model accuracy over the FaceForensics++ dataset using EfficientNet as our primary model. We also discuss several model reduction techniques such as architectural changes, model pruning and knowledge distillation to cut our model footprint by half while preserving much of our model’s original performance. Finally, we demonstrate how our proposed model can be turned into a deepfake detection application at the video level for solution completeness.