IJTIMOIY MUHANDISLIK HUJUMLARINI ANIQLASH VA ULARGA QARSHI KURASHISHNING INNOVATSION INTELLEKTUAL USULLARI VA ALGORITMLARINI ISHLAB CHIQISH
Annotatsiya
The article proposes a five-stage dialogue-level model for detecting social engineering attacks carried out through e-mail, SMS, messengers and voice calls. It was shown that existing tools analyse messages separately and therefore detect attacks that unfold as a multi-stage conversation too late. The model combines semantic features from a multilingual transformer, psychological persuasion principles and contextual features in a GRU layer, updates a risk score on a 0-100 scale after each new turn and warns the user before a harmful action occurs. Depending on the risk score, the conversation is passed, the user is warned or the conversation is blocked. Federated learning was used to train the model without centralising the confidential data of banks and mobile operators, and the SHAP method was used to produce understandable alerts. Proposals were made on organisational and technical countermeasures, evaluation of early detection performance and pilot testing at banks and operators.
Hali tarjima qilinmagan