ApaNet: adversarial perturbations alleviation network for face verification

Area of Science:

• Computer Vision
• Deep Learning
• Cybersecurity

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks, where subtle input perturbations can cause misclassifications.
• Face verification systems, critical for security, are susceptible to these adversarial manipulations, posing significant risks.

Purpose of the Study:

• To develop a robust defense mechanism against adversarial attacks specifically for face verification tasks.
• To introduce the adversarial perturbations alleviation network (ApaNet) for mitigating adversarial perturbations in facial images.

Main Methods:

• Implemented a novel network, ApaNet, utilizing stacked residual blocks to alleviate adversarial perturbations.
• Employed supervised learning on the Labeled Faces in the Wild (LFW) dataset, using legitimate and adversarial examples.
• Utilized middle and high-layer activations from FaceNet to define a loss function for optimizing ApaNet.

Main Results:

• Empirical results on LFW, YouTube Faces DB, and CASIA-FaceV5 datasets demonstrate ApaNet's effectiveness against white-box and black-box adversarial attacks.
• ApaNet significantly outperforms existing defense techniques in protecting face verification systems.

Conclusions:

• The proposed ApaNet provides a strong defense against adversarial attacks in face verification.
• This research contributes to enhancing the security and reliability of deep learning models in critical applications like facial recognition.