With the rapid advancement of the Internet and the proliferation of Internet of Things (IoT) devices, the diversity of malicious URLs has increased. These changes are reflected not only in their increasing number but also in the escalating technological complexity, which poses significant security threats to IoT networks. Traditional detection methods relying on manual rules or specific patterns are insufficient to address these challenges. There is a pressing need for more sophisticated URL detection techniques to enhance the accuracy and comprehensiveness of malicious URL identification, especially in IoT environments. This study integrates the feature purification network FP-net into malicious URL classification and introduces the FCBA model utilizing CNN-BiLSTM-Attention as the feature extractor for FP-net. The model employs CNN for effective local feature extraction, BiLSTM for capturing sequence information, and incorporates the attention mechanism. Experimental findings indicate that the FCBA model achieves an accuracy of 94.66% in malicious URL classification, surpassing CNN-BiLSTM-Attention by 0.65%. This demonstrates that FP-net enhances the performance of the CNN-BiLSTM-Attention model, and the FCBA model excels in URL classification tasks.