A new privacy attack network for remote sensing images classification with small training samples.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Overfitting remains a significant challenge for privacy attacks on small-sample remote sensing data.
• Existing methods struggle to effectively mitigate noise and inherent attributes within training data.

Purpose of the Study:

• To propose a novel deep learning network, the Joint Residual Network (JRN), for privacy object classification in small-sample remote sensing images.
• To address and solve the overfitting problem in privacy attack scenarios with limited data.

Main Methods:

• Introduced the Joint Residual Network (JRN), which fuses bottom and top feature maps using matrix joint, differing from standard residual network addition.
• Integrated JRN into the GoogleNet model, creating the GoogleNet-Feat model for experiments.
• Conducted benchmark experiments on UCMLU and WHU-RS datasets to evaluate performance.

Main Results:

• The GoogleNet-Feat model, incorporating JRN, demonstrated superior classification accuracy on small-sample datasets.
• On the UCMLU dataset, GoogleNet-Feat achieved 1.66% higher accuracy than GoogleNet and 1.87% higher than GoogleNet-R.
• On the WHU-RS dataset, GoogleNet-Feat showed a 1.04% increase over GoogleNet and a 3.12% increase over GoogleNet-R.

Conclusions:

• The Joint Residual Network (JRN) effectively reduces the extraction of training set noise and inherent attributes by convolution layers.
• GoogleNet-Feat, utilizing JRN, achieves the highest classification accuracy, successfully mitigating overfitting issues in small-sample remote sensing data privacy attacks.