The recent global spread of Monkeypox (Mpox), particularly the emergence of a highly transmissible strain in 2024, poses a significant public health threat beyond endemic regions. However, timely diagnosis remains challenging, especially in resource-limited settings where PCR testing is slow and costly. Deep Learning (DL)-based skin lesion classification offers a promising diagnostic alternative, yet most existing approaches rely on computationally expensive Transfer Learning (TL) models that often lack explainability. In this study, we benchmark eleven state-of-the-art (SOTA) TL architectures under a unified protocol using both feature and parameter transfer strategies across four studies. We then propose MicroMPoxNet, a novel lightweight model integrating a Separable Convolution-based Residual Attention Network (SCRAN) with Squeeze and Excitation (SE) blocks, optimized for memory efficiency, fast inference, and clinical interpretability on edge devices. We introduce Mpox2025, a new Mpox skin lesion dataset, and evaluate all models across four studies: feature transfer (Study 0), parameter transfer with binary classification without augmentation (Study 1), with augmentation (Study 2), and multiclass classification on a public dataset (Study 3). MicroMPoxNet achieved 99.99% accuracy with an AUC of 1.00 in binary classification (Study 2) and 99.12% accuracy with an AUC of 1.00 in multiclass classification (Study 3). With only 0.63 million parameters, an 86% reduction relative to SOTA models, it achieves the lowest computational complexity. Post hoc explainability methods, including Grad CAM, LIME, and Integrated Gradients, confirm that predictions consistently focus on clinically relevant lesion regions.