Memristor Based Neuromorphic Adaptive Resonance Theory for One-Shot Online Learning and Network Intrusion Detection
Memristor Based Neuromorphic Adaptive Resonance Theory for One-Shot Online Learning and Network Intrusion Detection
复制标题
DOI：
10.1145/3407197.3407608
复制
发表时间：
2020-07
期刊：
影响因子：
--
通讯作者：
Md. Shahanur Alam;C. Yakopcic;G. Subramanyam;T. Taha
中科院分区：
文献类型：
--
作者：
Md. Shahanur Alam;C. Yakopcic;G. Subramanyam;T. Taha
随着计算机网络变得更加先进，以极高的效率进行可靠的入侵检测的必要性大大增加。因此，在这项工作中，我们提出了一种能够在线学习的网络入侵检测的一次性学习系统。自适应谐振理论在定制的基于低功耗忆阻器的神经形态硬件中实现。该系统能够与现有知识区分开来，进行增量学习。赢家通吃电路由一个电容器和cmos计时电路实现，它可以找到获胜的神经元，并只控制获胜神经元的权重更新。找到一个获胜神经元所需的时间被确定在纳秒范围内。在以前已知的数据集和零日数据集上对该系统的性能进行了评估。使用零日报文的检测准确率为99.97%，对于已知攻击的检测准确率为99.99%。此外，使用不同的警戒参数和学习率对系统进行了测试。此外，还研究了电容器端阈值电压的变化对学习和检测精度的影响。
As computer networks become more advanced, the necessity for reliable intrusion detection at extreme efficiency has vastly increased. Thus, in this work we present a one shot learning system capable of online learning for network intrusion detection. Adaptive resonance theory is implemented in custom low power memristor-based neuromorphic hardware. The system is capable of discriminating with existing knowledge to learn incrementally. The winner take all circuitry is implemented with a capacitor and CMOS timing circuit that finds the winning neuron and controls the weight update for only the winning neuron. The time required to find a winning neuron was determined to be in the nanosecond range. The performance of the system was evaluated on both previously known and zero-day datasets. The detection accuracy using zero-day packets is 99.97%, and 99.99% for the known attacks. Furthermore, the system was tested using various vigilance parameters and learning rates. The variation of threshold voltage across the capacitor was also investigated to observe the effect on learning and detection accuracy.