Abstract: Online social network (OSN), as core platform for information exchange, currently face serious intrusion threats. However, existing OSN intrusion detection techniques exhibit poor detection performance when dealing with issues such as high dimensionality, diverse datasets with different types of structures, significant semantic differences, and mismatched dynamic features. Therefore, an intrusion detection method based on residual timeattention with feature selection (RTAS) is proposed. The method utilizes the pretrained language model BERT for data preprocessing and designs a classifier based on residual timeattention. The model effectively captures contextual features in a wide range of text information through a bidirectional LSTM and attention mechanism. Meanwhile, an adaptive feature selection method based on deep reinforcement learning is proposed, which utilizes adaptive learning to obtain the optimal feature set. The experiment shows that the proposed method achieves accuracies of 98.53%, 98.68%, and 98.33% in detecting multiple threat patterns on datasets from Facebook, Google+, and Twitter, respectively. The average accuracy on the three datasets exceeds other mainstream methods.

Key words: deep reinforcement learning, residual attention, feature selection, online social network, intrusion detection

摘要： 在线社交网络(online social network, OSN)作为信息交互的核心平台，当前面临严重的入侵威胁.现有OSN入侵检测技术在面对高维度、多样性数据集中不同类型结构、语义差异巨大以及动态特征失配等问题时检测性能较差.因此提出一种基于特征选择和时间残差注意力(residual timeattention with feature selection, RTAS)的入侵检测方法.利用预训练语言模型BERT(bidirectional encoder representations form transormers)进行数据预处理，并设计了一种基于时间残差注意力的分类器，模型通过双向长短期记忆网络(long shortterm memory, LSTM)和注意力机制有效捕获文本信息中的上下文特征.同时，提出了一种基于深度强化学习的自适应特征选择方法，该方法利用自适应学习获得最优的特征集合.实验表明，在Facebook,Google+，Twitter的数据集上，该方法在检测多种威胁模式方面分别达到了98.53%,98.68%,98.33%的准确率，平均准确率较现有主流方法有明显提升.

关键词: 深度强化学习, 残差注意力, 特征选择, 在线社交网络, 入侵检测