Defense Against Adversarial Attacks by Reconstructing Images.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are susceptible to adversarial examples, which are images with subtle, imperceptible modifications designed to cause misclassification.
• Existing defense mechanisms often struggle to maintain performance on clean data while effectively mitigating adversarial attacks.

Purpose of the Study:

• To propose a novel image reconstruction network for defending CNNs against adversarial attacks.
• To develop a robust defense that reconstructs adversarial examples into clean images.

Main Methods:

• Utilizing a residual block structure for precise mapping from adversarial to clean images.
• Employing perceptual loss to minimize error amplification during reconstruction.
• Integrating randomization layers to further suppress noise and enhance resilience against iterative attacks.

Main Results:

• The proposed network significantly reduces the impact of adversarial perturbations.
• The defense method shows minimal influence on the prediction accuracy of clean images.
• The model demonstrates superior performance compared to existing model-agnostic defenses during inference.

Conclusions:

• The developed image reconstruction network offers an effective defense against adversarial attacks on CNNs.
• The method exhibits strong generalization capabilities and can be combined with other defense strategies.