Layer-wise Orthogonal Training (LOT)

This repo provides the implementation of the paper "LOT: Layer-wise Orthogonal Training on Improving $\ell_2$ Certified Robustness". The code is improved based on the code repo of SOC.

Prerequisites

Follow the instructions in SOC to set up the environment.

Training an LOT network

python train_robust.py --conv-layer lot --activation ACT --block-size BLOCKS --dataset DATASET --gamma GAMMA --opt-level O0 --residual

• ACT: maxmin or hh1.
• BLOCKS: 1, 2, 3, 4, 5, 6, 7, 8
• DATASET: cifar10/cifar100.
• GAMMA: certificate regularization coefficient
• The LOT does not support O2 optimization as it requires a high float number precision.
• Use --lln to enable last layer normalization