ABCAttack: A Gradient-Free Optimization Black-Box Attack for Fooling Deep Image Classifiers

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial perturbations, leading to classification errors.
• Existing adversarial attack methods often require gradient evaluation or substitute model training, limiting their efficiency.

Purpose of the Study:

• To propose an efficient adversarial attack model for DNNs.
• To generate adversarial samples without gradient evaluation or substitute model training.
• To enhance the likelihood of classification task failure due to adversarial perturbation.

Main Methods:

• Developed an adversarial attack model utilizing the Artificial Bee Colony (ABC) algorithm.
• Generated adversarial samples to probe DNN vulnerabilities.
• Evaluated the attack's success rate in a black-box setting across MNIST, CIFAR-10, and ImageNet datasets.

Main Results:

• Achieved high attack success rates: 100% on MNIST, 98.6% on CIFAR-10, and 90.00% on ImageNet.
• Demonstrated effectiveness with fewer queries in black-box scenarios.
• Successfully bypassed existing defense mechanisms against adversarial attacks.

Conclusions:

• The proposed ABCAttack is an effective method for generating adversarial samples against DNNs.
• The attack is model-agnostic, applicable to various model structures and sizes.
• This research opens new avenues for developing robust deep learning evasion attacks and defenses.