ELAA: An Ensemble-Learning-Based Adversarial Attack Targeting Image-Classification Model.

Area of Science:

• Artificial Intelligence Security
• Machine Learning Vulnerabilities

Background:

• White-box adversarial attacks require model specifics, limiting real-world application.
• Black-box attacks offer a practical alternative, with reinforcement learning (RL) showing promise for evasion policies.
• Existing RL-based attacks exhibit suboptimal success rates.

Purpose of the Study:

• To enhance black-box adversarial attack effectiveness against image classification models.
• To address the limitations of current reinforcement learning (RL) approaches in adversarial attacks.
• To develop a more robust and successful adversarial attack strategy.

Main Methods:

• Proposing an ensemble-learning-based adversarial attack (ELAA).
• Aggregating and optimizing multiple reinforcement learning (RL) base learners.
• Evaluating ELAA's performance against image classification models.

Main Results:

• The ensemble model achieved a 35% higher attack success rate compared to single models.
• ELAA demonstrated a 15% improvement in attack success rate over baseline methods.
• The research highlights vulnerabilities in learning-based image classification models.

Conclusions:

• Ensemble learning enhances the efficacy of reinforcement learning-based adversarial attacks.
• ELAA offers a more practical and successful approach to black-box adversarial attacks.
• The findings underscore the need for robust defenses against sophisticated AI security threats.