Improving adversarial robustness by a coupling rejection strategy

Last update: Jan 06, 2023

Related tags

Deep Learning Rectified-Rejection

Overview

Adversarial Training with Rectified Rejection

The code for the paper Adversarial Training with Rectified Rejection.

Environment settings and libraries we used in our experiments

This project is tested under the following environment settings:

OS: Ubuntu 18.04.4
GPU: Geforce 2080 Ti or Tesla P100
Cuda: 10.1, Cudnn: v7.6
Python: 3.6
PyTorch: >= 1.6.0
Torchvision: >= 0.6.0

Acknowledgement

The codes are modifed based on Rice et al. 2020, and the model architectures are implemented by pytorch-cifar.

Training Commands

Below we provide running commands training the models with the RR module, taking the setting of PGD-AT + RR (ResNet-18) as an example:

python train_cifar.py --model_name PreActResNet18_twobranch_DenseV1 --attack pgd --lr-schedule piecewise \
                                              --epochs 110 --epsilon 8 \
                                              --attack-iters 10 --pgd-alpha 2 \
                                              --fname auto \
                                              --batch-size 128 \
                                              --adaptivetrain --adaptivetrainlambda 1.0 \
                                              --weight_decay 5e-4 \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate \
                                              --dataset 'CIFAR-10' \
                                              --ATframework 'PGDAT' \
                                              --SGconfidenceW

The FLAG --model_name can be PreActResNet18_twobranch_DenseV1 (ResNet-18) or WideResNet_twobranch_DenseV1 (WRN-34-10). For alternating different AT frameworks, we can set the FLAG --ATframework to be one of PGDAT, TRADES, CCAT.

Evaluation Commands

Below we provide running commands for evaluations.

Evaluating under the PGD attacks

The trained model is saved at trained_models/model_path, where the specific name of model_path is automatically generated during training. The command for evaluating under PGD attacks is:

python eval_cifar.py --model_name PreActResNet18_twobranch_DenseV1 --evalset test --norm l_inf --epsilon 8 \
                                              --attack-iters 1000 --pgd-alpha 2 \
                                              --fname trained_models/model_path \
                                              --load_epoch -1 \
                                              --dataset 'CIFAR-10' \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate

Evaluating under the adaptive CW attacks

The parameter FLAGs --binary_search_steps, --CW_iter, --CW_confidence can be changed, where --detectmetric indicates the rejector that needs to be adaptively evaded.

python eval_cifar_CW.py --model_name PreActResNet18_twobranch_DenseV1 --evalset adaptiveCWtest \
                                              --fname trained_models/model_path \
                                              --load_epoch -1 --seed 2020 \
                                              --binary_search_steps 9 --CW_iter 100 --CW_confidence 0 \
                                              --threatmodel linf --reportmodel linf \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate \
                                              --detectmetric 'RR' \
                                              --dataset 'CIFAR-10'

Evaluating under multi-target and GAMA attacks

The running command for evaluating under multi-target attacks is activated by the FLAG --evalonMultitarget as:

python eval_cifar.py --model_name PreActResNet18_twobranch_DenseV1 --evalset test --norm l_inf --epsilon 8 \
                                              --attack-iters 100 --pgd-alpha 2 \
                                              --fname trained_models/model_path \
                                              --load_epoch -1 \
                                              --dataset 'CIFAR-10' \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate \
                                              --evalonMultitarget --restarts 1

The running command for evaluating under GAMA attacks is activated by the FLAG --evalonGAMA_PGD or --evalonGAMA_FW as:

python eval_cifar.py --model_name PreActResNet18_twobranch_DenseV1 --evalset test --norm l_inf --epsilon 8 \
                                              --attack-iters 100 --pgd-alpha 2 \
                                              --fname trained_models/model_path \
                                              --load_epoch -1 \
                                              --dataset 'CIFAR-10' \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate \
                                              --evalonGAMA_FW

Evaluating under CIFAR-10-C

The running command for evaluating on common corruptions in CIFAR-10-C is:

python eval_cifar_CIFAR10-C.py --model_name PreActResNet18_twobranch_DenseV1 \
                                              --fname trained_models/model_path \
                                              --load_epoch -1 \
                                              --dataset 'CIFAR-10' \
                                              --twobranch --useBN \
                                              --selfreweightCalibrate

Improving adversarial robustness by a coupling rejection strategy

Related tags

Overview

Adversarial Training with Rectified Rejection

Environment settings and libraries we used in our experiments

Acknowledgement

Training Commands

Evaluation Commands

Evaluating under the PGD attacks

Evaluating under the adaptive CW attacks

Evaluating under multi-target and GAMA attacks

Evaluating under CIFAR-10-C

Owner

Tianyu Pang

Code for "Causal autoregressive flows" - AISTATS, 2021

Iterative Training: Finding Binary Weight Deep Neural Networks with Layer Binarization

This is the second place solution for : UmojaHack Africa 2022: African Snake Antivenom Binding Challenge

Library extending Jupyter notebooks to integrate with Apache TinkerPop and RDF SPARQL.

A collection of loss functions for medical image segmentation

The official code for PRIMER: Pyramid-based Masked Sentence Pre-training for Multi-document Summarization

SHIFT15M: multiobjective large-scale fashion dataset with distributional shifts

Explore the Expression: Facial Expression Generation using Auxiliary Classifier Generative Adversarial Network

Waymo motion prediction challenge 2021: 3rd place solution

This is the implementation of "SELF SUPERVISED REPRESENTATION LEARNING WITH DEEP CLUSTERING FOR ACOUSTIC UNIT DISCOVERY FROM RAW SPEECH" submitted to ICASSP 2022

SurfEmb (CVPR 2022) - SurfEmb: Dense and Continuous Correspondence Distributions

MIM: MIM Installs OpenMMLab Packages

A series of Python scripts to access measurements from Fluke 28X meters. Fluke IR Remote Interface required.

Code of Periodic Activation Functions Induce Stationarity

Model of an AI powered sign language interpreter.

code for "AttentiveNAS Improving Neural Architecture Search via Attentive Sampling"

An official implementation of the paper Exploring Sequence Feature Alignment for Domain Adaptive Detection Transformers

Official code for Next Check-ins Prediction via History and Friendship on Location-Based Social Networks (MDM 2018)

Code from Daniel Lemire, A Better Alternative to Piecewise Linear Time Series Segmentation

Point Cloud Denoising input segmentation output raw point-cloud valid/clear fog rain de-noised Abstract Lidar sensors are frequently used in environme