Beetle Antennae Search: Using Biomimetic Foraging Behaviour of Beetles to Fool a Well-Trained Neuro-Intelligent...

Area of Science:

• Computer Science
• Artificial Intelligence
• Neuroscience

Background:

• Deep Convolutional Neural Networks (CNNs) are advanced AI models for image classification, inspired by human brain structures.
• Simple organisms possess basic survival instincts that manifest as intelligent behavior, despite limited cognitive abilities.

Purpose of the Study:

• To demonstrate that a simple algorithm, mimicking beetle behavior, can effectively fool CNNs in image classification.
• To propose a computationally efficient adversarial attack method.

Main Methods:

• Developed a novel algorithm inspired by the behavior of a single beetle.
• Applied the algorithm to perturb single pixels in images for classification tasks.
• Evaluated performance on LeNet-5 and ResNet architectures using the CIFAR-10 dataset.

Main Results:

• The proposed beetle-inspired algorithm achieved a high success rate in fooling CNNs.
• The method is computationally efficient compared to other adversarial attack algorithms.
• The algorithm requires fewer search particles than swarm-based metaheuristic approaches.

Conclusions:

• A single-pixel perturbation, guided by simple biological behavior, can compromise CNN image classification.
• The findings raise significant concerns regarding the robustness and security of AI systems.
• This research highlights the need for more resilient AI models against sophisticated adversarial attacks.