An accurate pixel-Level explainable approach for CNNs and its application.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are prevalent in image classification.
• Interpreting the decision-making process of trained CNN models remains a challenge.
• Existing explainable methods often lack sufficient interpretation accuracy.

Purpose of the Study:

• To develop a novel pixel-level explainable approach for CNNs.
• To address the limitations of insufficient interpretation accuracy in current methods.
• To design a scheme for enhancing the adversarial robustness of CNN models.

Main Methods:

• A novel pixel-level explainable approach was developed and tested.
• Experiments were conducted on PyTorch team published CNN models.
• A scheme to enhance adversarial robustness was designed based on the explainable approach.

Main Results:

• The presented approach achieved 100% accuracy in interpreting the pixel-level classification basis of input images.
• The designed scheme effectively improved the adversarial robustness of CNN models.
• The technique demonstrated transferability across different CNN model structures.

Conclusions:

• The novel pixel-level explainable approach significantly improves interpretation accuracy for CNNs.
• The developed method offers an effective strategy for enhancing CNN adversarial robustness.
• This approach provides a valuable tool for understanding and securing CNN models.