Monkeypox, a re-emerging zoonotic disease, has become a growing global public health concern due to its rapid transmission and high visual similarity to other dermatological conditions such as chickenpox and measles. This resemblance complicates early clinical diagnoses, particularly in resource-limited settings where laboratory testing capabilities are scarce. Deep learning methods capable of detecting monkeypox from skin images offer a promising alternative to manual inspection and PCR-based diagnoses. However, existing approaches are often limited by poor dataset quality, weak generalization ability, and insufficient model interpretability. To address these challenges, this paper proposes an attention-enhanced hybrid deep learning framework for the automated classification of monkeypox skin lesions. Specifically, the proposed model employs DenseNet-121 and EfficientNet-B4 as parallel convolutional feature extractors and integrates Convolutional Block Attention Modules (CBAM) to adaptively emphasize lesion-related features while suppressing background interference. Experimental results demonstrate that the proposed framework outperforms conventional deep learning baselines, thereby achieving an accuracy exceeding 91% and a Cohen's Kappa score above 0.88 on the primary dataset. Furthermore, the model exhibits a strong generalization capability, thereby maintaining classification accuracies above 90% across multiple external public datasets. To enhance the transparency and clinical reliability, explainable artificial intelligence (XAI) methods are employed to visualize the model's decision-making process. In addition, quantitative interpretability metrics are used to assess the reliability and consistency of the generated explanations, thus highlighting the model's potential for practical clinical applications.