【摘要】：針對大規(guī)模網(wǎng)絡(luò)態(tài)勢要素獲取時間復(fù)雜度較高和攻擊樣本不平衡導(dǎo)致小類樣本分類精度不高的問題,提出一種基于深度自編碼網(wǎng)絡(luò)的態(tài)勢要素獲取機(jī)制。在該機(jī)制下,利用優(yōu)化后的深度自編碼網(wǎng)絡(luò)作為基分類器,識別數(shù)據(jù)類型。一方面,在自編碼網(wǎng)絡(luò)的逐層訓(xùn)練中,提出一種結(jié)合交叉熵(CE)函數(shù)和反向傳播(BP)算法的訓(xùn)練規(guī)則,克服傳統(tǒng)的方差代價函數(shù)更新權(quán)值過慢的缺陷;另一方面,在深度網(wǎng)絡(luò)的微調(diào)和分類階段,提出一種主動在線采樣(AOS)算法應(yīng)用于分類器中,通過在線選擇用于更新網(wǎng)絡(luò)權(quán)值的攻擊樣本,達(dá)到總樣本的去冗余和平衡各類攻擊樣本數(shù)量的目的,從而提高小類攻擊樣本的分類精度。經(jīng)對實(shí)例數(shù)據(jù)的仿真分析,該方案有較好的態(tài)勢要素獲取精度,并能有效減少數(shù)據(jù)傳輸時的通信開銷。
[Abstract]:Aiming at the problem of high time complexity of acquisition of situation elements in large-scale networks and unbalance of attack samples resulting in low classification accuracy of small class samples, a novel approach based on deep self-coding network is proposed. In this mechanism, the optimized depth self-coding network is used as the base classifier to identify the data types. On the one hand, a new training rule combining cross-entropy (CE) function and back-propagation (BP) algorithm is proposed to overcome the disadvantage of the traditional variance cost function updating the weight too slowly in the self-coding network layer by layer training; on the other hand, In the stage of fine-tuning and classification of depth network, an active on-line sampling (AOS) algorithm is proposed to be applied to classifier, and attack samples are selected online to update network weights. The goal of eliminating redundancy of total samples and balancing the number of attack samples is achieved so as to improve the classification accuracy of small attack samples. The simulation results show that the scheme has better precision of acquisition of situation elements and can effectively reduce the communication cost of data transmission.
【作者單位】： 重慶市移動通信重點(diǎn)實(shí)驗(yàn)室(重慶郵電大學(xué));
【基金】：國家自然科學(xué)基金資助項(xiàng)目(61271260,61301122) 重慶市科委自然科學(xué)基金資助項(xiàng)目(cstc2015jcyjA40050)~~
【分類號】：TP393.08
，

本文編號：2148917