【作者】 张澎；

【导师】 赵欢； 刘晓雁；

【作者基本信息】 湖南大学 ， 计算机科学与技术， 2009， 硕士

【摘要】 入侵检测是近十年发展起来的一种动态监测、预防或抵御系统入侵行为的安全机制。目前入侵检测有许多模型和方法,而神经网络和模式识别等技术的引入使网络安全的智能检测研究成为热点。神经网络具有自学习、自适应的能力,只要提供系统的审计数据或网络数据包,神经网络就可以通过自学习从中提取正常的用户或系统活动特征模式,并检测出异常活动的攻击模式。这些特性使其在入侵检测中得到了很好的应用。目前最流行的神经网络学习算法是BP算法,BP学习算法是基于梯度下降这一本质,不可避免地会带来以下缺点:学习过程收敛速度慢;算法不完备,容易陷入局部极值,当学习速率设置过高时,可能产生振荡;鲁棒性差,网络性能对网络的初始设置比较敏感。这使得基于BP神经网络的入侵检测存在高漏报率和误报率,本文针对入侵检测的效率问题和准确性问题,提出了一种基于量子遗传算法优化BP神经网络的入侵检测模型,该模型基于量子遗传算法的全局搜索和BP网络局部精确搜索的特性,将量子遗传算法和BP算法有机结合,利用量子遗传算法优化BP神经网络的权重和阈值。针对基本量子遗传算法存在容易早熟和局部搜索能力弱等缺陷,提出了改进的量子遗传算法,并对量子遗传算法的各个环节进行了细致的分析与重新设计,包括量子比特编码、量子门更新、量子变异等。实验证明运用此方法可在一定程度上提高入侵检测的效率和准确性。更多还原

【Abstract】 Intrusion detection is developed as a dynamic monitoring system to prevent or resist the intrusion of the security mechanism in the past decade.There are many intrusion detection models and methods, while the neural network and pattern recognition technologies enable the introduction of intelligent intrusion detection to become a hot research. Neural network has self-learning, adaptive capacity, as long as the system’s audit data or network packets, neural networks can through self-study to extract the normal characteristics of the user or system activity patterns and detect unusual activity patterns of attack. These features make it applicate in intrusion detection very good.The most popular neural network learning algorithm is BP algorithm.However, it has the following drawbacks as the gradient descent property of the BP algorithm: the convergence speed is slow; the algorithm is incomplete and tends to get the local extreme; the algorithm can be oscillatory when the learning ratio is set too high; the algorithm is not robust and the network performance is sensive to the initial configuration of the network. All of these drawbacks can lead to the high omission rate and false alarm rate when BP algorithm is applied in instrusion detection. This paper presented an instrusion detection model based on quantum genetic algorithm and BP neural network, which can deal these phenomenoms propertly. The model takes advantage of the global search property of the quantum genetic algorithm and the exact local search characteristics of the BP network. The weight and the thresholds of the BP neural network is optimized by the quantum genetic algorithm.As basic quantum genetic algorithm has weeklocal search ability and is easy to premature, this paper also proposed an improved quantum genetic algorithm. Many aspects of the algorithm have been re-designed with considerately analysis including the quantum bit encoding, fitness function design, update quantum gates, quantum variations, etc.Experiments show that this method can be used to improve the efficiency and accuracy of intrusion detection.更多还原