An Efficient TFRecord-based Data Pipeline for Deepfake Detection Training

Key Points

• This research compares different data pipelines for deepfake detection training to improve overall efficiency and performance.
• Compared Baseline, DiskCache, and TFRecord pipelines using datasets of varying sizes (500,000, 250,000, and 50,000 images).
• Trained MobileNetV2 model under identical conditions for performance metrics.
• Analyzed accuracy, speed, and stability of each pipeline.
• TFRecord pipeline demonstrated the best balance of accuracy, speed, and stability.
• TFRecord proved to be the most practical solution for deepfake detection training.
• Highlights the importance of data pipeline design in machine learning applications.

Abstract