MsMemoryGAN: A Multiscale Memory GAN for Palm-Vein Adversarial Purification.

Area of Science:

• Biometrics
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel in vein recognition but are vulnerable to adversarial attacks.
• Adversarial attacks introduce subtle input perturbations, causing DNNs to misclassify vein patterns.
• Existing defense mechanisms often fall short in robustly handling these attacks.

Purpose of the Study:

• To propose MsMemoryGAN, a novel defense model for filtering adversarial perturbations in vein recognition.
• To enhance the robustness of vein recognition systems against sophisticated adversarial attacks.

Main Methods:

• Designed a multiscale memory autoencoder (MsMemoryAE) with a memory module (MM) for learning normal vein patterns at various scales.
• Introduced a learnable similarity metric within the MM (LSMM) to purify input features by retrieving relevant normal patterns.
• Integrated pixel, perceptual, and adversarial losses for high-quality image reconstruction and perturbation removal.

Main Results:

• MsMemoryGAN successfully reconstructs normal vein patterns while failing to reconstruct adversarial perturbations.
• Extensive experiments on public datasets demonstrate effective removal of diverse adversarial perturbations.
• The proposed defense significantly improves vein recognition accuracy when faced with adversarial samples.

Conclusions:

• MsMemoryGAN offers a robust defense against adversarial attacks in vein recognition tasks.
• The memory module with learnable similarity is key to purifying adversarial features.
• This approach substantially enhances the reliability and accuracy of biometric systems.