Defense-optimized BERT architecture against digital arrest attacks in intelligent systems

Digital arrest (DA) is a type of spear phishing attack that is recently used by the scammers to extract money from the victims. In DA, the scammer impersonated as the law enforcement officer and gave instructions to the victim through emails, SMS, or telephone. As the scammers copy the format and pattern of official communication; hence these communications are not detected by the conventional phishing detection models, in this context, in this paper we proposed an AI-driven BERT for the detection of DA scam. The proposed model extracts the import information from the communication using multi-heading attention, channel attention, and spatial attention blocks. The proposed model employs multihead attention to extract semantic dependencies and contextual relationships between tokens, while the channel attention module dynamically recalibrates feature map importance to prioritize relevant dimensions. In addition, the incorporation of spatial attention enhances the location of critical DA cues by refining the spatial representation of features. Due to the use of dedicated attention extraction blocks, the proposed model achieved superior performance with an accuracy of 97.8% and an F1 score of 98.3%, significantly outperforming conventional models such as GRU, LSTM and RNN. Furthermore, the proposed model requires 99.7% fewer FLOPs, has 99.6% fewer parameters, and demonstrates a time efficiency improvement of up to 87.8% compared to these traditional models.