gatheluck / PaperReading

Notes about papers (in Japanese)
0 stars 0 forks source link

[2017] SafetyNet: Detecting and Rejecting Adversarial Examples Robustly #146

Open gatheluck opened 4 years ago

gatheluck commented 4 years ago

論文リンク

公開日(yyyy/mm/dd)

2017/04/01 ICCV 2017

概要

TeX

% 2017/04/01
@inproceedings{
    lu2017safetynet,
    title={SafetyNet: Detecting and Rejecting Adversarial Examples Robustly},
    author={Jiajun Lu and Theerasit Issaranon and David Forsyth},
    booktitle=ICCV,
    year={2017}
}
gatheluck commented 4 years ago

ネットワークの最終層の活性化を多段階に量子化したものをRBF-SVMに入力して、敵対的サンプルを検出する。「On Detecting Adversarial Perturbations」と似ているが勾配情報の推定を難しくすることで、頑健になっている。