Adversarial-Aware Deep Learning System Based on a Secondary Classical Machine Learning Verification Approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models excel in image classification but are vulnerable to adversarial attacks.
• Adversarial attacks exploit the neural network structures inherent in deep learning models.
• Classical machine learning models, like random forest, lack neural network designs, suggesting potential immunity.

Purpose of the Study:

• To investigate the vulnerability of classical machine learning models to adversarial attacks.
• To propose a novel adversarial-aware deep learning system for enhanced image classification security.
• To evaluate the effectiveness of the proposed system against state-of-the-art adversarial defense methods.

Main Methods:

• Experimental evaluation of classical machine learning models against popular adversarial attacks.
• Development of a hybrid deep learning system incorporating a classical machine learning model as a secondary verification layer.
• Testing the proposed system on the CIFAR-100 dataset.

Main Results:

• Classical machine learning models demonstrated immunity to adversarial attacks, supporting the initial hypothesis.
• The proposed adversarial-aware deep learning system effectively detected adversarial attacks through output mismatch.
• The hybrid system achieved superior performance compared to existing state-of-the-art adversarial defense systems.

Conclusions:

• Classical machine learning models offer a robust defense against adversarial attacks due to their non-neural network architecture.
• Integrating classical models as verification systems in deep learning enhances adversarial robustness without compromising primary model accuracy.
• The proposed adversarial-aware system presents a promising direction for secure and reliable image classification.