Detecting anti-forensic deepfakes with identity-aware multi-branch networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake detection systems excel at identifying conventional forgeries but struggle with adversarial samples designed to evade detection.
• Adversarial samples use subtle perturbations to mask forgery artifacts, leading traditional binary classifiers to misclassify them as authentic.
• Existing methods are vulnerable to anti-forensic techniques that specifically target and deceive detection algorithms.

Purpose of the Study:

• To develop a robust deepfake detection framework capable of identifying both conventional and anti-forensic (adversarial) deepfakes.
• To improve the resilience of deepfake detection systems against deliberate evasion strategies.
• To enhance the classification accuracy for real, forged, and anti-forensic media.

Main Methods:

• Proposed a multi-channel feature extraction framework for deepfake detection.
• Employed a three-class classification strategy to differentiate between real, forged, and anti-forensic samples.
• Utilized identity-preserving facial representations, spatial, and frequency domain features for comprehensive analysis.

Main Results:

• The proposed method demonstrated significant improvements in robustness against anti-forensic attacks.
• Maintained high accuracy in detecting conventional deepfakes.
• Effectively distinguished between real, forged, and adversarial deepfake samples in experimental datasets.

Conclusions:

• The multi-channel feature extraction and three-class classification approach offers enhanced robustness against adversarial deepfakes.
• This framework provides a more reliable solution for deepfake detection in the presence of sophisticated evasion techniques.
• The findings contribute to advancing the field of digital forensics and media authenticity verification.