Publications

Detailed Information

Robustness of deep neural networks to adversarial attack : 심층신경망의 적대적 공격에 대한 강건성: 휴리스틱 방법론부터 검증가능한 방법론까지
from heuristic methods to certified methods

DC Field Value Language
dc.contributor.advisor이재욱-
dc.contributor.author이성윤-
dc.date.accessioned2022-04-20T07:51:02Z-
dc.date.available2022-04-20T07:51:02Z-
dc.date.issued2021-
dc.identifier.other000000166880-
dc.identifier.urihttps://hdl.handle.net/10371/178936-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000166880ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 자연과학대학 수리과학부, 2021.8. 이재욱.-
dc.description.abstractDeep learning has shown successful results in many applications. However, it has been demonstrated that deep neural networks are vulnerable to small but adversarially designed perturbations in the input which can fool the neural network. There have been many studies on such adversarial attacks and defenses against them. However, Athalye et al. [1] have shown that most defenses rely on specific predefined adversarial attacks and can be completely broken by stronger adaptive attacks. Thus, certified methods are proposed to guarantee stable prediction of input within a perturbation set. We present this transition from heuristic defense to certified defense, and investigate key features of certified defenses, tightness and smoothness.-
dc.description.abstract딥러닝은 다양한 분야에서 성공적인 성능를 보여주고 있다. 그러나 심층신경망은 적대적 공격이라 불리우는, 입력값에 작은 섭동을 주어 신경망을 사용자가 원치 않는 방향으로 행동하도록 하는 공격에 취약하다. 적대적 공격의 발견 이후로, 다양한 적대적 공격과 이에 대한 방어 방법론과 관련하여 많은 연구들이 진행되었다. 그러나 Athalye et al. [1] 에서 대부분의 기존 방어 방법론들이 특정 적대적 공격만을 가정하고 설계되어 더 강한 적응가능한 적대적 공격에 의해 공격 가능하다는 문제점이 밝혀졌다. 따라서 입력값에 대해 섭동가능한 영역내에서 안정적인 행동을 보증할 수 있는 검증가능한 방법론이 제안되어왔다. 본 학위 논문에서는, 휴리스틱 방법론과 검증가능한 방법론에 대해 알아보고, 검증가능한 방법론에서 중요한 요소인 상한의 밀착성과 목적함수의 매끄러움에 대해서 분석한다.-
dc.description.tableofcontents1 Introduction 1
2 Heuristic Defense 3
2.1 Heuristic Defense 3
2.1.1 Background 3
2.2 Gradient diversity regularization 5
2.2.1 Randomized neural network 5
2.2.2 Expectation over Transformation (EOT) 5
2.2.3 GradDiv 6
2.2.4 Experiments 11
3 Certified Defense 21
3.1 Certified Defense 21
3.1.1 Background 21
3.2 Tightness of the upper bound 24
3.2.1 Lipschitz-certifiable training with tight outer bound 24
3.2.2 Experiments 31
3.3 Smoothness of the objective 36
3.3.1 Background 36
3.3.2 What factors influence the performance of certifiable training? 39
3.3.3 Tightness and smoothness 46
3.3.4 Experiments 47
4 Conclusion and Open Problems 58
Appendix A Appendix for 2.2 60
A.1 Experimental Settings 60
A.1.1 Network Architectures 60
A.1.2 Batch-size, Training Epoch, Learning rate decay,Warmup, and Ramp-up periods 61
A.2 Variants of GradDiv-mean (2.2.17) 61
A.3 Additional Results on "Effects of GradDiv during Training" 61
A.4 Additional Results on Table 2.1 62
A.5 In the case of n > 20 in Figure 2.7 62
A.6 RSE [48] as a baseline 62
Appendix B Appendix for 3.2 68
B.1 The proof of the proposition 3.1.1 68
B.2 Outer Bound Propagation 69
B.2.1 Intuition behind BCP 69
B.2.2 Power iteration algorithm 69
B.2.3 The circumscribed box $out_\infty(h^{(k+1)}(\mathbb{B}^{(k)}_2))$ 71
B.2.4 BCP through residual layers 71
B.2.5 Complexity Analysis 72
B.3 Experimental Settings 72
B.3.1 Data Description 72
B.3.2 Hyper-parameters 73
B.3.3 Network architectures 73
B.3.4 Additional Experiments 74
Appendix C Appendix for 3.3 81
C.1 Experimental Settings 81
C.1.1 Settings in Section 3.3.2 82
C.1.2 Settings in Table 3.4 83
C.2 Interval Bound Propagation (IBP) 84
C.3 Details on Linear Relaxation 84
C.3.1 Linear relaxation explained in CROWN [79] 84
C.3.2 Dual Optimization View 85
C.4 Learning curves for variants of CROWN-IBP 87
C.5 Mode Connectivity 87
C.6 ReLU 91
C.7 $\beta$- and $\kappa$-schedulings 91
C.8 one-step vs multi-step 92
C.9 Train with $\epsilon_{train}\geq\epsilon_{test}$ 92
C.9.1 $\epsilon_{train}\geq\epsilon_{test}$ on MNIST 92
C.9.2 $\epsilon_{train}=1.1\epsilon_{test}$ on CIFAR-10 93
C.10 Training time 94
C.11 Loss and Tightness violin plots 95
C.12 Comparison with CAP-IBP 95
C.13 ReLU Stability 95
Bibliography 103
Abstract (in Korean) 113
-
dc.format.extentiv, 113-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectDeep Learning-
dc.subjectAdversarial Robustness-
dc.subjectCertified Defense-
dc.subject딥러닝-
dc.subject적대적 강건성-
dc.subject검증가능한 방어-
dc.subject.ddc510-
dc.titleRobustness of deep neural networks to adversarial attack-
dc.title.alternative심층신경망의 적대적 공격에 대한 강건성: 휴리스틱 방법론부터 검증가능한 방법론까지-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorSungyoon Lee-
dc.contributor.department자연과학대학 수리과학부-
dc.description.degree박사-
dc.date.awarded2021-08-
dc.title.subtitlefrom heuristic methods to certified methods-
dc.identifier.uciI804:11032-000000166880-
dc.identifier.holdings000000000046▲000000000053▲000000166880▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share