Decision-based Batch Processing Attack Method

Deep Convolutional Neural Networks (CNNs) have significantly impacted various domains, ranging from text recognition to natural language processing, computer vision, and cross-modal sentiment analysis in the community in recent years. Nevertheless, adversarial samples can cause those models to make wrong predictions and pose a huge security threat to production systems and daily life. Studying adversarial attacks exposes the weaknesses of CNNs and highlights the associated security risks, guiding improvements in model robustness. We explored how deep neural networks are vulnerable to adversarial perturbations and highlighted the shortcomings of current computational paradigms. These paradigms are challenged by their inability to produce adversarial instances that are broadly applicable across extensive datasets, thereby limiting the robustness of the models. Despite advances in general adversarial perturbation and SPSA optimization, limitations like high computational demands and sensitivity to model architecture persist. We introduce an enhanced Dec Attack algorithm, incorporating batch accuracy loss and SPSA optimization techniques to tackle this challenge. This enhancement facilitates the efficient execution of adversarial stimuli in decision-based black-box scenarios. The algorithm’s efficacy is evidenced by its superior performance compared to alternative black-box approaches, thereby augmenting the practicality and viability of deploying adversarial strategies.