《電子技術(shù)應(yīng)用》
您所在的位置:首頁(yè) > 其他 > 設(shè)計(jì)應(yīng)用 > 可解釋的深度網(wǎng)絡(luò)抗噪音干擾性逐層評(píng)估方法
可解釋的深度網(wǎng)絡(luò)抗噪音干擾性逐層評(píng)估方法
網(wǎng)絡(luò)安全與數(shù)據(jù)治理
周瑞豐1,欒瑞鵬2,張超1
1.大連理工大學(xué)數(shù)學(xué)科學(xué)學(xué)院; 2. 91550部隊(duì)
摘要: 在復(fù)雜應(yīng)用場(chǎng)景中,深度神經(jīng)網(wǎng)絡(luò)易受噪音數(shù)據(jù)干擾,如何客觀、有效、可靠地評(píng)估深度網(wǎng)絡(luò)的抗噪音干擾性已成為智能化技術(shù)開(kāi)發(fā)的關(guān)鍵問(wèn)題之一。然而,現(xiàn)有的評(píng)估方法只能對(duì)網(wǎng)絡(luò)抗噪音干擾性進(jìn)行整體評(píng)估,且不具備可解釋的理論依據(jù)。為了解決上述問(wèn)題,基于閔可夫斯基差線性可分性度量,提出一種深度神經(jīng)網(wǎng)絡(luò)逐層抗噪音干擾性評(píng)估方法,實(shí)現(xiàn)了神經(jīng)網(wǎng)絡(luò)抗噪音干擾性的可解釋性評(píng)估。該方法可以對(duì)深度網(wǎng)絡(luò)各隱層數(shù)據(jù)映射行為進(jìn)行定量分析,解析噪音干擾對(duì)網(wǎng)絡(luò)各隱層的影響機(jī)理,進(jìn)而建立面向網(wǎng)絡(luò)隱層的可解釋抗噪音干擾性評(píng)估方法。
中圖分類(lèi)號(hào):TP181文獻(xiàn)標(biāo)識(shí)碼:ADOI:10.19358/j.issn.2097-1788.2025.09.007
引用格式:周瑞豐,欒瑞鵬,張超. 可解釋的深度網(wǎng)絡(luò)抗噪音干擾性逐層評(píng)估方法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,2025,44(9):46-50.
Explainable methods for layer-wise evaluating noise resistance of deep networks
Zhou Ruifeng1,Luan Ruipeng 2,Zhang Chao1
1. School of Mathematical Science, Dalian University of Technology; 2. Unit 91550
Abstract: In complex application scenarios, deep neural networks are susceptible to noise data interference. How to objectively, effectively, and reliably evaluate the noise resistance of deep networks has become one of the key issues in the development of intelligent technology. However, the existing evaluation methods can only evaluate the noise resistance of the whole network, and have no explainable theoretical basis. To address the above problems, based on Minkowski difference based linear separability measure, a layer-wise noise resistance evaluation method for deep neural networks is proposed, achieving an explainable evaluation of the network′s noise resistance performance. This method can quantitatively analyze the data mapping behavior of each hidden layer of the deep network, explain the influence mechanism of noise interference on each hidden layer of the network, and then establish an explainable noise resistance evaluation method for the network′s hidden layers.
Key words : noise resistance evaluation; explainability; neural network

引言

隨著神經(jīng)網(wǎng)絡(luò)在圖像分類(lèi)[1]、語(yǔ)義分割[2]等領(lǐng)域的長(zhǎng)足發(fā)展,其被廣泛地應(yīng)用于工業(yè)生產(chǎn)[3]、自動(dòng)駕駛[4]、醫(yī)學(xué)[5]等新的任務(wù)場(chǎng)景中。然而,這些新的任務(wù)場(chǎng)景往往更加復(fù)雜,其中包含著干擾輸入數(shù)據(jù)的多種噪音,會(huì)顯著影響網(wǎng)絡(luò)的性能。同時(shí),在這些任務(wù)場(chǎng)景中,網(wǎng)絡(luò)判斷出錯(cuò)所造成的后果往往要更加嚴(yán)重。因此,如何評(píng)估網(wǎng)絡(luò)對(duì)任務(wù)場(chǎng)景中噪音的抗干擾性,從而判斷網(wǎng)絡(luò)在該任務(wù)場(chǎng)景中的適用性,成為了一個(gè)需要解決的問(wèn)題。

目前,通過(guò)生成讓網(wǎng)絡(luò)判斷出錯(cuò)的最小擾動(dòng)的對(duì)抗樣本[6-8]或者生成任務(wù)場(chǎng)景[9-10]中常見(jiàn)的模擬噪音都可以作為生成噪音的方法,并進(jìn)一步得到網(wǎng)絡(luò)在面對(duì)噪音下的數(shù)值化抗干擾性指標(biāo)。但是,現(xiàn)有的網(wǎng)絡(luò)評(píng)估方法大多對(duì)網(wǎng)絡(luò)整體進(jìn)行評(píng)估,無(wú)法分析噪音對(duì)網(wǎng)絡(luò)輸出的影響機(jī)理,缺乏可解釋性,并且評(píng)估結(jié)果也無(wú)法對(duì)網(wǎng)絡(luò)改進(jìn)起到積極的指導(dǎo)作用。

為了解決上述問(wèn)題,本文引入基于閔可夫斯基差的線性可分性度量(Minkowski Difference based Linear Separability Measure, MDLSM)作為分類(lèi)神經(jīng)網(wǎng)絡(luò)的隱層分析工具,借助MDLSM對(duì)網(wǎng)絡(luò)隱層的數(shù)據(jù)映射行為進(jìn)行分析,得到噪音對(duì)網(wǎng)絡(luò)各個(gè)隱層的擾動(dòng),確定噪音對(duì)網(wǎng)絡(luò)的影響機(jī)理,從而實(shí)現(xiàn)網(wǎng)絡(luò)的可解釋性抗干擾性逐層評(píng)估。


本文詳細(xì)內(nèi)容請(qǐng)下載:

http://ihrv.cn/resource/share/2000006706


作者信息:

周瑞豐1,欒瑞鵬2,張超1

(1.大連理工大學(xué)數(shù)學(xué)科學(xué)學(xué)院,遼寧大連116024;

2. 91550部隊(duì),遼寧大連116023)


subscribe.jpg

此內(nèi)容為AET網(wǎng)站原創(chuàng),未經(jīng)授權(quán)禁止轉(zhuǎn)載。