As more traditional businesses, such as banking and finance, are transferred to online platforms or the cloud, the deepening of system interaction with users and the improvement of technology-based defence system make cyber attackers focus more on human beings, leading to serious financial consequences. This attack utilising social engineering often exploits human nature's weakness. Its complexity, language variability and inductivity are difficult to defend effectively. Therefore, this paper proposes a model for detecting social engineering attacks based on deep neural network by reviewing current methods for social engineering detection, in terms of phishing, deception and content-based detection, in addition to examining deep learning algorithms with excellent data performance. Through the processing and analysis of natural language in chat history, the attention-based Bi-LSTM is used to capture and mine the context semantics, and the ResNet is used to integrate user characteristics and content characteristics for classification and judgment. By describing the features of social engineering attacks and online conversations, the feasibility and effectiveness of the proposed model are demonstrated from the perspective of algorithm selection and applicability.