This repository contains code for the paper Certifying Confidence via Randomized Smoothing by Aounon Kumar, Alexander Levine, Soheil Feizi and Tom Goldstein. It is built on the publicly availabe code repository for the paper Certified Adversarial Robustness via Randomized Smoothing by (Cohen et al. 2019).
Randomized smoothing has been shown to provide good certified-robustness guarantees for high-dimensional classification problems. It uses the probabilities of predicting the top two most-likely classes around an input point under a smoothing distribution to generate a certified radius for a classifier's prediction. However, most smoothing methods do not give us any information about the confidence with which the underlying classifier (e.g., deep neural network) makes a prediction. This work proposes a method to generate certified radii for the prediction confidence of the smoothed classifier.
Follow the instructions listed by Cohen et al. 2019 in their GitHub repository under Getting started to set up the project. Obtain a copy of the pre-trained models from here and extract them into a directory.
To certify average prediction score of a ResNet-110 model on the CIFAR-10 dataset, run:
python code/certify.py cifar10 [path-to-models]/models/cifar10/resnet110/noise_0.25/checkpoint.pth.tar 0.25 log_files/cifar10/pred_score/noise_0.25 --skip 20 --batch 400
python code/certify.py cifar10 [path-to-models]/models/cifar10/resnet110/noise_0.50/checkpoint.pth.tar 0.50 log_files/cifar10/pred_score/noise_0.50 --skip 20 --batch 400
To repeat the same experiments with margin as the confidence measure, use:
python code/certify.py cifar10 [path-to-models]/models/cifar10/resnet110/noise_0.25/checkpoint.pth.tar 0.25 log_files/cifar10/margin/noise_0.25 --skip 20 --batch 400 --confidence_measure margin
python code/certify.py cifar10 [path-to-models]/models/cifar10/resnet110/noise_0.50/checkpoint.pth.tar 0.50 log_files/cifar10/margin/noise_0.50 --skip 20 --batch 400 --confidence_measure margin
To run the same experiments for ImageNet with a ResNet-50 model, download ImageNet and preprocess the val directory to look
like the train directory by running this script.
Set the IMAGENET_DIR environment variable to the location of the ImageNet dataset. Modify the above commands used for
the CIFAR-10 experiments as shown below.
export IMAGENET_DIR=[path-to-imagenet]
python code/certify.py imagenet [path-to-models]/models/imagenet/resnet50/noise_0.25/checkpoint.pth.tar 0.25 log_files/imagenet/pred_score/noise_0.25 --skip 100 --batch 400
python code/certify.py imagenet [path-to-models]/models/imagenet/resnet50/noise_0.50/checkpoint.pth.tar 0.50 log_files/imagenet/pred_score/noise_0.50 --skip 100 --batch 400
python code/certify.py imagenet [path-to-models]/models/imagenet/resnet50/noise_0.25/checkpoint.pth.tar 0.25 log_files/imagenet/margin/noise_0.25 --skip 100 --batch 400 --confidence_measure margin
python code/certify.py imagenet [path-to-models]/models/imagenet/resnet50/noise_0.50/checkpoint.pth.tar 0.50 log_files/imagenet/margin/noise_0.50 --skip 100 --batch 400 --confidence_measure margin
Finally, to compute certified accuracies use the file cert_acc.py:
python cert_acc.py [path to log file] [score to certify]
For example:
python cert_acc.py log_files/imagenet/pred_score/noise_0.25 0.5