Inspector for Face Forgery Detection: Defending Against Adversarial Attacks From Coarse to Fine

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Face forgery detection is crucial for social security due to rising concerns.
• Current detectors are vulnerable to adversarial perturbations, which can disrupt subtle forgery cues.
• Existing defenses often degrade natural accuracy by altering input data or detectors.

Purpose of the Study:

• To develop a novel adversarial defense paradigm for face forgery detectors.
• To enhance robustness against adversarial attacks without compromising natural accuracy.
• To introduce a defense mechanism that preserves both instances and detectors.

Main Methods:

• A novel defensive paradigm named Inspector was designed.
• Inspector employs a coarse-to-fine defense strategy.
• The coarse stage filters evident adversarial perturbations; the fine stage addresses imperceptible ones.

Main Results:

• Inspector achieves state-of-the-art performance against diverse adversarial perturbations.
• The method effectively preserves natural accuracy of face forgery detectors.
• Experiments validated performance across various datasets and detectors.

Conclusions:

• Inspector offers an effective solution for defending face forgery detectors against adversarial attacks.
• The coarse-to-fine approach successfully balances robustness and natural accuracy.
• This work provides a significant advancement in secure and reliable face forgery detection.