Phishing websites continue to deceive users by impersonating trustworthy websites. We propose a multimodal deep learning model that combines image features extracted from EfficientNetB2 and HTML text features through DistilBERT. These features are coupled using a Multi-Head Attention mechanism to exploit the relationship between patterns, thereby improving the accuracy of phishing website detection. When tested on a benchmark dataset, our model achieves an accuracy of 97.35%, significantly outperforming models that use only one type of data. This result demonstrates that combining image and text signals outperforms traditional methods that rely only on URLs to detect phishing websites.