基于击键序列特征提取的用户身份认证

【作者】 傅博；

【导师】 王晅；

【作者基本信息】 陕西师范大学 ， 信号与信息处理， 2008， 硕士

【摘要】 击键序列身份认证是保障计算机信息安全的一种重要手段。20世纪80年代,Gaines等人首次提出基于击键动力学(keystroke dynamics)特征进行身份认证。其实现简单,运行成本低廉,但突出的问题是如何有效地从击键序列中提取出关键信息,从而在贝叶斯统计算法的基础上,将基于击键动力学特征的判别问题抽象为三类模式识别问题,即入侵类、怀疑类和合法用户类,对怀疑类采用基于概率论的二次识别机制,同时利用统计学相关公式的推导,给出了该系统安全级别的选取范围及标准,并通过实验比较了两种有代表性的贝叶斯模型应用二分类和三分类判别法前后系统性能的差别。实验结果表明错误拒绝率(False reject rate,FRR)和错误通过率(False accept rate,FAR)都得到了有效的降低。但是识别过程中作为衡量模式集间差异的距离量度是否应该选取广泛采用的欧几里德距离?本文在接下来的研究中发现测地距(geodesic distance)能够更加有效地表达这一距离量度。同时受认知生理学关于记忆(神经流形)和视觉流形相关性研究的启发,提出记忆和敲击键盘的触觉认知流形也具备相关性,在此基础上引入流形学习算法。以测地距作为距离量度,给出了算法对于密码长度的理论适用范围,并通过实验证明该算法能够更加有效地降低FRR和FAR。更多还原

【Abstract】 Authentication based on keystroke pattern is an important method to safeguard computer information. In 1980’s, Gaines, etc. presented an authentication approach based on keystroke dynamics in the first time. This approach is easy to be realized and cheap to be applied. But a significant problem is how to efficiently extract key information from the keystroke sequence and then decrease error rate. For this purpose, researchers have been doing a lot of work and many approaches have been proposed: approaches based on neural networks, support vector machine, genetic algorithm, etc. However, these approaches more or less share the defects of long convergence time and large training set. It is for that reason, these approaches cannot be applied in real time detection. To conquer these defects, a new approach based on continuous statistical model has been presented. With the merit of small training set and fast convergence rate, it can successfully hurdle those shortcomings. But this kind of approach also processes a disadvantage: its error rate is relatively high and preprocessing is needed in order to get low error rate. Aim at this demerit, based on Bayesian statistical algorithm, this paper abstracts pattern recognition of keystroke dynamics to three-class recognition, that is, intrusion class, suspicious class and normal class. Then mechanism of reorganization is applied to the suspicious class. Moreover, the range and criteria of safe rate for the system is provided through statistical deduction. The difference between classic approach and the new approach is theoretically discussed. The experimental result shows that the false reject rate and false accept rate are both decreased.However, a question is should the Euclidean distance be a necessary measurement for difference between pattern set? In the latter research, geodesic distance has been found to be a more ideal measurement. Meanwhile, illuminated by the research that there exists connection between memory(neural manifold) and visual manifold, a similar connection between memory and haptic manifold is proposed. Based on these, manifold learning algorithm is introduced. The new approach applies geodesic distance and provides theoretical range of the length of key. Experimental result proves that the FRR and FAR are further decreased.更多还原