Arabic SMS classification is a difficult task because of dialect differences, non-standard spelling, code-switching, and the paucity of realistic annotated corpora. This paper introduces an effective hybrid architecture of deep learning model that combines the AraELECTRA transformer and a Bidirectional Long Short-Term Memory (BiLSTM) network in order to classify Arabic spam effectively. The system uses Arabic-oriented preprocessing, such as normalization, Farasa-based tokenization, stopwords removal, and noise filtering, to enhance the representation of short and informal texts. They were experimented with a human-validated Arabic translation of the UCI SMS Spam Collection and a diversity of new Arabic Iraqi data, and dialectal perturbation and spelling distortion tests to test the output of robustness in the presence of realistic noise. The fusion model that was proposed was better than standalone architectures with 98.8 and 95.3 accuracies. The computational analysis proved a lightweight design with smaller parameter size and more efficient inference, which is supported by the fact that it can be deployed in resource-limited mobile settings. The findings showed that a significant trade-off existed between performance and the cost of computation between efficient transformer embeddings and sequential contextual modeling. Irrespective of the limitations of the datasets, the present work has set a feasible ground in Arabic spam detection and has offered some information to the further literature that addresses dialect-enriched corpora and adversarial robustness testing.