ytsmiling / lmt

Public code for a paper "Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks."

LMT

Public code for a paper "Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks."

update: Pre-print and code were updated to version 2 (2018/5/23).

How to train:

python3 train.py (configuration-file-name).py

what is the configuration file?

example:

python3 train.py config/parseval_svhn/default.py --gpu 0

How to evaluate with attacks:

python3 evaluate.py (result-dir-of-trained-network) (attack-configuration).py

This script calculates each value in inequality (5) in Sec.6 and create inequality-(attack_name).npy under a specified result directory.

what is the result directory?

example:

python3 evaluate.py result/config/parseval_svhn/default-00 config/attack/cw_100_1.py

Reference

Y. Tsuzuku, I. Sato, M. Sugiyama: Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks, (2018), url, bibtex