Introduction

Code to the paper "Interpretability based Neural Network Repair".

The implementation of this project is based on pytorch 1.13 with python 3.7.0

INNER was evaluated on CIFAR10, CIFAR100, GTSRB, and ImageNet10. CIFAR10 and CIFAR100 can be downloaded automatically (in ~/datasets by default). GTSRB and ImageNet10 need to be placed in the dataset folder. I will upload ImageNet10 soon.

Modify get_standard and get_backdoor functions in ~/utils/makedataset.py if you want to train models on your own data.

Useage

put your dataset, e.g., cifar10 in the folder of './dataset/cifar10', so the code can find the data

1. Model generation

Train models with certain architecture and dataset, specify the certain repair scenario and it will be trained as well:

python main.py --mode train --arch vgg13_dense --set cifar10 --save_dir ./checkpoints/cifar10 --rep_type bd

This will generate folders containing the model weights, other operations depend on the args to determine the path of weights.

2. Train model probes

You must specify the number of samples and target task to train the probe

python main.py --mode probe  --probe_train_num 1000 --arch vgg13_dense --set cifar10 --rep_type bd --save_dir ./checkpoints/cifar10

3. repair model

You can specify the number of samples and layers for repair

python main.py --mode rep --probe_train_num 1000 --arch vgg13_dense --set cifar10 --rep_type bd --rep_layer_num 3 --repair_sample_num 1000 --save_dir ./checkpoints/cifar10 --rep_dir ./checkpoints/repaired/

4. test repair

Test will generate a log file containing the results

python  main.py --mode test  --probe_train_num 1000 --arch vgg13_dense --set cifar10 --rep_type bd --rep_layer_num 3 --repair_sample_num 1000 --save_dir ./checkpoints/cifar10 --rep_dir ./checkpoints/repaired/