A Distributed Black-Box Adversarial Attack Based on Multi-Group Particle Swarm Optimization

Area of Science:

• Deep Learning Security
• Artificial Intelligence (AI) Vulnerabilities
• Cybersecurity

Background:

• Adversarial attacks in deep learning pose risks due to their subtlety and potential real-world impact.
• Existing black-box attacks often require numerous queries, increasing detection likelihood by AI systems.
• Previous gradient-free methods using Particle Swarm Optimization (PSO) suffered from low success rates due to local optima entrapment.

Purpose of the Study:

• To develop a more effective black-box adversarial attack method that minimizes query counts.
• To enhance the success rate of PSO-based adversarial attacks by addressing the local optima problem.
• To demonstrate the attack's efficacy and scalability through distributed execution and real-world testing.

Main Methods:

• Implemented a distributed attack strategy across multiple nodes to reduce query volume and detection probability.
• Utilized Multi-Group PSO with Random Redistribution (MGRR-PSO) for improved perturbation generation and local optima avoidance.
• Introduced MGRR-PSO for efficient perturbation pruning, outperforming standard iterative methods.
• Conducted experiments on MNIST, CIFAR-10, ImageNet, and a real-world attack against Google Cloud Vision.

Main Results:

• Achieved a 100% attack success rate on MNIST and CIFAR-10 datasets.
• Successfully fooled Google Cloud Vision, demonstrating real-world applicability.
• Maintained a significantly lower query count compared to existing methods.
• Showcased improved scalability and robustness through distributed execution.

Conclusions:

• The proposed distributed MGRR-PSO-based adversarial attack effectively increases success rates while minimizing queries.
• This method offers a scalable and stealthier alternative for black-box adversarial attacks.
• The attack's real-world success against Google Cloud Vision validates its practical implications for AI security.