jin's blog

  • 홈
  • 태그
  • 미디어로그
  • 위치로그
  • 방명록

They Are Features 1

[Adversarial Robustness] Adversarial Examples Are Not Bugs, They Are Features

NeurIPS 2019에 spotlight으로 발표된 논문이다. Adversarial example은 non robust feature로 부터 발생되는 것임을 보이며, 즉 모델의 버그가 아니고 data의 내재적인 특성으로 인해 발생되는 현상이다 라고 주장하는 논문임 DNN이 adversarial example에 취약하다는 것은 잘 알려져있다. 이전 연구들은 이런 원인이 input space의 high-dims에서 오거나 train data의 통계적인 fluctuation으로부터 발생한다거나, 또는 우리가 사용하는 relu의 positive부분에서 local linearity가 발생해서 공격이 쉽게 당한다. 등 다양한 주장이 있었다. 또한 이러한 attack 을 방어하는 여러 전략들을 다뤄져왔다. 이 논문..

IT/Paper 2021.11.03
이전
1
다음
더보기
프로필사진

jin's blog

Endure

  • 분류 전체보기 (50)
    • IT (44)
      • Paper (23)
      • Reinforcement Learni.. (0)
      • Probability (0)
      • Deep learning (6)
      • Spark (5)
      • Python (4)
      • Computer vision (4)
      • Data Structure (1)
    • 관심사 (2)
      • 낚시 (0)
      • 피아노 (2)
      • 일상 (0)

Tag

They Are Features, Interpretability Beyond Feature Attribution:Quantitative Testing with Concept Activation Vectors, smoothGrad, Learning Directed Exploration Strategies, R-CNN, Fast R-CNN, TCAV, Concept vector, Deconvolution Network, intergrated gradient, XAI, Never Give Up, RL논문, Regularizing Trajectory Optimization with Denoising Autoencoders, Paper리뷰, CAV, Adversarial Examples Are Not Bugs, vision transformer, Quantifying Attention Flow in Transformers, Axiomatic Attribution for Deep Networks,

최근글과 인기글

  • 최근글
  • 인기글

최근댓글

공지사항

페이스북 트위터 플러그인

  • Facebook
  • Twitter

Archives

Calendar

«   2025/05   »
일 월 화 수 목 금 토
1 2 3
4 5 6 7 8 9 10
11 12 13 14 15 16 17
18 19 20 21 22 23 24
25 26 27 28 29 30 31

방문자수Total

  • Today :
  • Yesterday :

Copyright © Kakao Corp. All rights reserved.

티스토리툴바