Surface defect detection is a critical process for ensuring the quality of printed circuit boards (PCB). In recent years, numerous convolutional neural networks have been proposed for this task. Among them, the cost-sensitive residual network (CS-ResNet) has demonstrated exceptional performance on imbalanced datasets by incorporating a cost-sensitive adjustment layer into the standard ResNet architecture. However, CS-ResNet still faces challenges in global representation learning and struggles to accurately capture tiny defect features. Additionally, its limited ability to model complex nonlinear relationships further constrains its performance. To address these issues, in this study, we propose the Kolmogorov-Arnold Attention Network (KA2Net). In KA2Net, we first introduce the attention mechanism via the Convolutional Block Attention Module (CBAM) to prioritize key defect regions, enabling the model to effectively detect tiny defects. Subsequently, we replace the traditional fully connected layer in the backbone network with Kolmogorov-Arnold Networks (KAN) to further enhance its accuracy and interpretability. Extensive experiments on a real-world PCB surface defect dataset demonstrated the superior performance of KA2Net.