Local artifacts amplification for deepfakes augmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of AI-generated content (AIGC) makes distinguishing real from forged facial images increasingly difficult.
• Existing deepfake detection methods often overlook the nuanced utilization of local artifacts.
• Traditional data augmentation techniques show limited efficacy for forgery detection, necessitating specialized approaches.

Purpose of the Study:

• To develop an efficient system for detecting forged facial images.
• To propose a novel augmentation method, Local Artifacts Amplification for Deepfakes Augmentation (LAADFA), to enhance deepfake detection.
• To improve the utilization of local artifacts and incorporate facial feature priors into detection models.

Main Methods:

• Proposed Local Artifacts Amplification for Deepfakes Augmentation (LAADFA) to amplify subtle artifacts in forged regions.
• Incorporated prior knowledge of similar facial feature patterns within defined facial regions.
• Aggregated results across all facial regions to enhance overall model performance.

Main Results:

• Achieved an Area Under the Curve (AUC) of 93.40% and an accuracy (Acc) of 87.03% on the WildDeepfake dataset.
• Demonstrated superior performance compared to traditional augmentation methods in intra-dataset evaluations.
• Showcased strong generalization capabilities through cross-dataset evaluations.

Conclusions:

• The proposed LAADFA method significantly improves deepfake detection accuracy and robustness.
• Effective amplification of local artifacts and incorporation of facial priors enhance model performance.
• The method exhibits strong generalization, making it suitable for diverse deepfake detection scenarios.