From Messages to Malicious Links: Joint Spam and Phishing Detection via Cross-Attentive Multimodal Learning

* A multilingual Bangla-English-Hindi spam-text corpus containing 11,260 messages and a merged phishing-URL corpus containing 1,248,283 URLs were assembled from six public sources. Five multilingual transformer models and seven classical and deep learning URL classifiers were benchmarked on these datasets as single-task baselines.

* A balanced hybrid dataset of 50,000 message-URL pairs was constructed, covering all four possible spam/phishing label combinations. Five hybrid architectures were then trained by combining a text encoder (MuRIL, mBERT, or DistilBERT) with a character-level URL encoder (GRU, CNN, or BiLSTM) using the same training protocol.

* A hybrid model combining MuRIL and a character-level GRU through multi-head cross-attention was proposed. In this architecture, the text representation directly queries the URL sequence. The proposed model achieved the best overall joint spam-text and phishing-URL performance among the five hybrid configurations evaluated.

* The proposed model was compared with recent spam and phishing detection studies to assess the benefits and limitations of joint text-URL fusion. The comparison showed that hybrid fusion improved performance in several cases, while single-task models remained better in some specialized tasks, including one case where the hybrid URL detection head performed worse than the best standalone URL classifier.