Online Alternate Generator against Adversarial Attacks.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Deep convolutional neural networks have driven progress in computer vision.
• Deep learning models are vulnerable to adversarial examples, which are images with imperceptible noise.
• Current defense methods are often inefficient and ineffective against unknown adversarial attacks.

Purpose of the Study:

• To propose a portable and efficient defense method against adversarial examples.
• To develop a defense that does not require access to or modification of target network parameters.

Main Methods:

• An online alternate generator synthesizes a new image from scratch for each input.
• The generator and synthesized image are updated alternately during inference.
• This approach avoids retraining attacked networks or augmenting training data.

Main Results:

• The proposed online alternate generator demonstrates superior performance compared to state-of-the-art defense models.
• The method effectively defends against gray-box adversarial attacks.
• The defense is portable and does not rely on specific network architectures.

Conclusions:

• The online alternate generator offers a robust and efficient defense against adversarial attacks in computer vision.
• This method provides a promising direction for enhancing the security of deep learning models against sophisticated threats.