Jointly Defending DeepFake Manipulation and Adversarial Attack Using Decoy Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Generative Adversarial Networks (GANs) enable realistic image and video synthesis.
• DeepFakes and adversarial attacks pose significant threats by manipulating visual content and misleading AI systems.
• Combined threats of DeepFakes and adversarial perturbations complicate defense strategies.

Purpose of the Study:

• To develop a novel deceptive mechanism for detecting DeepFake manipulations and adversarial attacks.
• To enhance the robustness of detection methods against combined threats.
• To propose a statistical hypothesis testing framework for identifying manipulated media.

Main Methods:

• Designed a deceptive model with two isolated sub-networks to generate specific two-dimensional random variables.
• Utilized a maximum likelihood loss function for training the deceptive model.
• Proposed a novel hypothesis testing scheme for detection using the trained deceptive model.

Main Results:

• The proposed deceptive mechanism demonstrated effectiveness in detecting DeepFake images and videos.
• The method successfully identified manipulated content across various unseen and compressed manipulation techniques.
• The statistical hypothesis testing approach proved robust against combined DeepFake and adversarial attacks.

Conclusions:

• The novel decoy mechanism offers a promising defense against sophisticated DeepFake and adversarial attacks.
• The statistical hypothesis testing framework provides a generalized solution for detecting manipulated visual media.
• This research contributes to securing digital media integrity in the face of advanced AI-driven threats.