关键词: ：Ｗｅｂ攻击检测, 文本预处理, 多头注意力机制, 聚焦处理

Abstract: A deep learning model is used to detect Web attacks and full HTTP texts are input to make the vocabulary larger, which causes model parameter overloads and increases storage costs. In addition, location uncertainty and semantic complexity of the attack payloads lead to a higher missing alarm rate. To solve the problems of model parameter overloads and missing attack payloads, this paper proposes a Web attack detection method based on the lightweight vocabulary cooperative memory focus processing model. Firstly, this novel method generates a lightweight vocabulary.Secondly, in combination with the preprocessing rules of the lightweight vocabulary, it preprocesses the HTTP texts according to the preprocessing rules likes aving, replacement, addition and discarding to reduce parameter overloads. Finally, this method uses a memory focus processing model based on bidirectional long and short term memory and the multi-head attention mechanism, which improves the memory ability and the focus processing ability of the attack loads to reduce the missing alarm rate. In the Simulation Dataset, the accuracy rate of this novel method is 98.66%, which is 3.19% higher than that of URL_WORD+GRU. Among the detected attack types, the lowest missing alarm rate is 0.60%. The experimental results demonstrate that the novel method can effectively alleviateparameter overloads, improve the detection accuracy and reduce the missing alarm rate.