Abstract:
The rapid advancement of deepfake technologies has emerged as a critical threat to the authenticity of digital media through the creation of highly realistic but fake visual content. The use of advanced video manipulation tools undermines public trust, poses security risks and challenges the integrity of information across digital platforms. Although significant progress has been made in deepfake image detection, research on identifying advanced video manipulation, particularly in low-resolution videos remains limited. To address this gap, this study propose a deepfake video detection framework by using transfer learning approach. Initially, our method analyzes each extracted frame using a CNN-based architecture to generate frame-level predictions, which are then aggregated by averaging. Based on a predefined threshold, the framework finally classifies the video as real or manipulated. For experimentation, we utilized a publicly available dataset named “FaceForensics++” containing a total of 2,000 real and manipulated videos of varying quality and resolution. We explored various CNN architectures including Xception, Densenet121, Inception ResNet V2, ResNet50, EfficientNet B3 along with rigorous hyperparameter tuning. Among these, the Xception architecture outperformed others by achieving a test accuracy of 94.5%. This research offers an effective solution to the growing challenge of deepfake detection that facilitates the development of robust and scalable tools that are vital for preserving information integrity in the industry 4.0.
