摘要
门限循环单元是一种循环神经网络模型的变体,它改进了长短期记忆模型。经验表明,循环神经网络被广泛应用于不同类型的机器学习场景中,在解决自然语言处理、语音识别、文本分类等问题均良中有好表现。通常情况下,循环神经网络模型会使用Softmax函数作为顶层分类器,用交叉熵函数计算损失。在判断网络数据是否为恶意入侵数据这项任务中,可以使用二分类器节省计算开销。因此,对传统模型予以改进,用线性支持向量机取代Softmax函数作为模型的顶层分类器,用基于边界的损失函数取代交叉熵函数。实验结果表明,改进的模型在提升准确率和时间复杂度方面比传统模型表现好。
Gated recurrent unit is a variant of recurrent neural network model,and it has improved long short-term memory model. With empirical evidence,recurrent neural network models are being applied to a wide range of machine learning tasks,and are proven to be effective in natural language processing,speech recognition,text classification and so on. Conventionally,the both models choose the Softmax function as their final output layer for prediction and the cross-entropy function to compute the loss. As for the task of judging whether the network data is intrusive,we can use binary-classifier to save computational cost. Therefore,we improve the conventional model as follows,using support vector machine in place of Softmax function as top layer classifier,using margin-based function in place of cross-entropy function. Results show that the improved model performs relatively better than conventional model in accuracy and time complexity.
作者
王禹
丁箐
罗弦
Wang Yu;Ding Qing;Luo Xian(School of Software, University of Science and Technology of China, Hefei 230000, Chin)
出处
《信息技术与网络安全》
2018年第4期30-33,共4页
Information Technology and Network Security
关键词
机器学习
循环神经网络
支持向量机
人工智能
入侵检测
machine learning
recurrent neural networks
support vector machine
artificial intelligence
intrusion detection
