Adversarial Attack Robustness of CNN Models for Arabic Traffic Panel Text Detection

The rapid progress of Artificial Intelligence (AI), notably in the field of convolutional neural networks (CNNs), has led to notable improvements in natural scene text detection. Despite these advances, CNN-based models remain susceptible to adversarial attacks that intentionally manipulate predictions, thereby compromising the robustness of AI-based systems. this vulnerability remains inadequately studied in the specific domain of Arabic traffic panel text detection, where high visual diversity and the intrinsic complexity of the Arabic script introduce additional challenges.This study examines the adversarial robustness of three commonly adopted CNN-based text detection architectures—CTPN, EAST, and TextBoxes++—when exposed to multiple attack scenarios. The evaluation is carried out using two Arabic traffic panel datasets, namely ASAYAR and STP, under varying attack configurations and perturbation intensities. To our knowledge, this work constitutes the first systematic comparison of the robustness of these models within this specific application context. The experimental findings highlight significant security weaknesses, underscoring the necessity for more robust and dependable text detection solutions. Furthermore, the outcomes contribute to guiding the design of resilient defense strategies for intelligent transportation systems.