DeepCover: Uncover the Truth Behind AI

DeepCover explains image classifiers using statistical fault localization and causal theory.

Videos: ECCV2020, ICCV2021

Install and Setup

Create a clean Docker container with Ubuntu 20.04

docker run -v ${PWD}:/home -it ubuntu:20.04

Commands

apt-get update && apt-get install pip git ffmpeg libsm6 libxext6 && pip install matplotlib seaborn tensorflow==2.3.0 keras==2.4.3 numpy==1.18.0 scipy==1.4.1 opencv-python && cd home && git clone https://github.com/theyoucheng/deepcover

Hello DeepCover

python ./src/deepcover.py --help
usage: deepcover.py [-h] [--model MODEL] [--inputs DIR] [--outputs DIR]
                    [--measures  [...]] [--measure MEASURE] [--mnist-dataset]
                    [--normalized-input] [--cifar10-dataset] [--grayscale]
                    [--vgg16-model] [--inception-v3-model] [--xception-model]
                    [--mobilenet-model] [--attack] [--text-only]
                    [--input-rows INT] [--input-cols INT]
                    [--input-channels INT] [--x-verbosity INT]
                    [--top-classes INT] [--adversarial-ub FLOAT]
                    [--adversarial-lb FLOAT] [--masking-value INT]
                    [--testgen-factor FLOAT] [--testgen-size INT]
                    [--testgen-iterations INT] [--causal] [--wsol FILE]
                    [--occlusion FILE]

To start running the Statistical Fault Localization (SFL) based explaining:

python ./src/deepcover.py --mobilenet-model --inputs data/panda --outputs outs --testgen-size 200

--mobilenet-model pre-trained keras model

--inputs input images folder

--outputs output images folder

--testgen-size the number of input mutants to generate for explaining (by default, it is 2,000)

More options

python src/deepcover.py --mobilenet-model --inputs data/panda/ --outputs outs --measures tarantula zoltar --x-verbosity 1 --masking-value 0

--measures to specify the SFL measures for explaining: tarantula, zoltar, ochiai, wong-ii

--x-verbosity to control the verbosity level of the explanation results

--masking-value to control the masking color for mutating the input image

To start running the causal theory based explaining:

python ./sfl-src/sfl.py --mobilenet-model --inputs data/panda --outputs outs --causal --testgen-iterations 50

--causal to trigger the causal explanation

--testgen-iterations number of individual causal refinement calls; by default, it’s 1

To load your own model

python src/deepcover.py --model models/gtsrb_backdoor.h5 --input-rows 32 --input-cols 32 --inputs data/gtsrb/ --outputs outs

--input-rows row number for the input image

--input-cols column number for the input image

Publications

@inproceedings{sck2021,
  AUTHOR    = { Sun, Youcheng
                and Chockler, Hana
                and Kroening, Daniel },
  TITLE     = { Explanations for Occluded Images },
  BOOKTITLE = { International Conference on Computer Vision (ICCV) },
  PUBLISHER = { IEEE },
  PAGES     = { 1234--1243 },
  YEAR = { 2021 }
}

@inproceedings{schk2020,
AUTHOR = { Sun, Youcheng
and Chockler, Hana
and Huang, Xiaowei
and Kroening, Daniel},
TITLE = {Explaining Image Classifiers using Statistical Fault Localization},
BOOKTITLE = {European Conference on Computer Vision (ECCV)},
YEAR = { 2020 }
}

Miscellaneous

Roaming Panda Dataset

Photo Bombing Dataset

DeepCover Site