Adversarial Defense Method Based on Latent Representation Guidance for Remote Sensing Image Scene Classification.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Deep neural networks (DNNs) excel in remote sensing image analysis but are vulnerable to adversarial examples.
• Adversarial attacks pose risks to regional and production safety, necessitating robust defense mechanisms.
• Existing defense methods often struggle against unknown adversarial noise.

Purpose of the Study:

• To propose an adversarial denoising method for remote sensing image scene classification.
• To enhance the robustness of DNNs against adversarial noise.
• To improve the safety and reliability of remote sensing image analyses.

Main Methods:

• A variational autoencoder (VAE) is trained for data reconstruction using clean datasets.
• At test time, normalized mutual information guides selective image reconstruction.
• Latent representation is iteratively updated via reconstruction loss to eliminate adversarial noise.

Main Results:

• The proposed method demonstrates effectiveness in remote sensing image scene classification.
• It achieves improved robust accuracy compared to state-of-the-art adversarial defense methods.
• The denoiser, trained solely on clean data, shows robustness against unknown adversarial noise.

Conclusions:

• The latent representation guidance adversarial denoising method is effective for remote sensing images.
• The approach enhances robustness against adversarial attacks, crucial for safety-critical applications.
• This method offers a promising direction for secure and reliable deep learning in remote sensing.