Improving one-pixel adversarial attacks using hybrid quantum-inspired differential evolution

One-pixel adversarial attacks expose critical vulnerabilities in deep neural networks by inducing misclassification through minimal input perturbations, posing serious risks for security-sensitive applications. Despite their practical significance, the effectiveness and efficiency of existing evolutionary optimization strategies for such attacks remain insufficiently understood. This study presents a systematic comparative evaluation of Differential Evolution (DE) and a hybrid Quantum-Inspired Evolutionary Algorithm combined with DE (QIEA + DE), which leverages quantum-inspired probabilistic search to enhance global exploration in black-box settings. Experiments are conducted on the CIFAR-10 and CIFAR-100 datasets using three convolutional neural network architectures: SimpleCNN, ResNet-32, and a VGG-inspired model. The methods are evaluated on 500 randomly selected test images per model using success rate, confidence drop, query complexity, and computational cost as performance metrics. Results show that Hybrid QIEA + DE achieves up to 7–13% higher average confidence reduction compared to DE, while incurring 1.6×–1.9× higher computational cost, highlighting a clear trade-off between attack effectiveness and efficiency. These findings provide practical guidance for adversarial evaluation and robustness assessment in resource-constrained environments.