Media Forensics Considerations on DeepFake Detection with Hand-Crafted Features

Area of Science:

• Media Forensics
• Computer Vision
• Artificial Intelligence

Background:

• DeepFake technology poses a significant threat to the integrity of video evidence.
• Current DeepFake detection methods predominantly rely on neural networks to learn distinguishing features.
• The interpretability of learned features in neural networks remains a challenge for validating detection decisions.

Purpose of the Study:

• To propose and evaluate an alternative approach to DeepFake detection using hand-crafted features.
• To investigate the interpretability and plausibility validation benefits of expert-defined features.
• To assess the detection performance and generalization capabilities of hand-crafted features compared to learned features.

Main Methods:

• Development and implementation of three distinct sets of hand-crafted features by domain experts.
• Application of three different feature fusion strategies to combine the hand-crafted features.
• Rigorous testing on three established reference databases to evaluate detection performance.

Main Results:

• Achieved peak Area Under the Curve (AUC) exceeding 0.95, comparable to state-of-the-art neural network methods.
• Demonstrated superior or similar generalization performance across different training and testing datasets compared to neural network approaches.
• Explored a data-centric examination approach for enhancing the maturity of forensic process modeling.

Conclusions:

• Hand-crafted features offer a viable and interpretable alternative for DeepFake detection.
• The proposed method demonstrates competitive accuracy and improved generalization, bolstering trust in forensic analysis.
• This research contributes to advancing media forensics through explainable AI techniques and robust detection strategies.