trustai/deepconcolic

Concolic Testing for Deep Neural Networks

DeepConcolic (Testing for Deep Neural Networks)

The tutorial document for this repository is in (tutorial.pdf)

General Introduction

This repository includes a few software packages, all of which are dedicated for the analysis of deep neural netowrks (or tree ensembles) over its safety and/or security properties.

1. DeepConcolic, a coverage-guided testing tool for convolutional neural networks. Now, it includes a major upgrade based on Bayesian Network based Abstraction.
2. testRNN, a coverage-guided testing tool for Long short-term memory models (LSTMs). LSTMs are a major class of recurrent neural networks.
3. EKiML, a tool for backdoor embedding and detection for tree ensembles.
4. GUAP: a generalised universal adversarial perturbation. It generates universersal adversarial perburbation that may be applied to many inputs at the same time.

In the following, after the installation and download of example models, we will present them one by one.

Installation

First of all, please set up a conda environment

conda create --name deepconcolic python==3.7
conda activate deepconcolic

This should be followed by installing software dependencies:

conda install opencv nltk matplotlib
conda install -c pytorch torchvision
pip3 install numpy==1.19.5 scipy==1.4.1 tensorflow\>=2.4 pomegranate==0.14 scikit-learn scikit-image pulp keract np_utils adversarial-robustness-toolbox parse tabulate pysmt saxpy keras menpo patool z3-solver pyvis

Download Example Models

We use Fashion-MNIST dataset as the running example. The following are two pre-trained mmodels, one larger and one smaller.

wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/small_model_fashion_mnist.h5
wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/large_model_fashion_mnist.h5

Tool 1 -- DeepConcolic: Concolic Testing for Convolutional Neural Networks

Concolic testing alternates between CONCrete program execution and symbOLIC analysis to explore the execution paths of a software program and to increase code coverage. In this paper, we develop the first concolic testing approach for Deep Neural Networks (DNNs). More specifically, we utilise quantified linear arithmetic over rationals to express test requirements that have been studied in the literature, and then develop a coherent method to perform concolic testing with the aim of better coverage. Our experimental results show the effectiveness of the concolic testing approach in both achieving high coverage and finding adversarial examples.

The paper is available at https://arxiv.org/abs/1805.00089.

In the following, we first present the original ASE2018 version, and then introduce two new upgrades (fuzzing engine and Bayesian network based abstraction).

ASE2018 Version

Work Flow

Sample Results

Command to Run

usage: python3 -m deepconcolic.main [-h] --dataset
                                    {OpenML:har,cifar10,fashion_mnist,mnist}
                                    --model MODEL --outputs DIR --criterion
                                    {nc,ssc,ssclp,bfc,bfdc} --norm {l0,linf}
                                    [--setup-only] [--init INT]
                                    [--max-iterations INT] [--save-all-tests]
                                    [--rng-seed SEED]
                                    [--extra-tests DIR [DIR ...]]
                                    [--filters {LOF}] [--norm-factor FLOAT]
                                    [--lb-hard FLOAT] [--lb-noise FLOAT]
                                    [--mcdc-cond-ratio FLOAT]
                                    [--top-classes CLS]
                                    [--layers LAYER [LAYER ...]]
                                    [--feature-index INT] [--dbnc-spec SPEC]
                                    [--dbnc-abstr PKL]

Concolic testing for Neural Networks

optional arguments:
  -h, --help            show this help message and exit
  --dataset {OpenML:har,cifar10,fashion_mnist,mnist}
                        selected dataset
  --model MODEL         the input neural network model (.h5 file or "vgg16")
  --outputs DIR         the output test data directory
  --criterion {nc,ssc,ssclp,bfc,bfdc}
                        the test criterion
  --norm {l0,linf}      the norm metric
  --setup-only          only setup the coverage critierion and analyzer, and
                        terminate before engine initialization and startup
  --init INT            number of test samples to initialize the engine
  --max-iterations INT  maximum number of engine iterations (use < 0 for
                        unlimited)
  --save-all-tests      save all generated tests in output directory; only
                        adversarial examples are kept by default
  --rng-seed SEED       Integer seed for initializing the internal random
                        number generator, and therefore get some(what)
                        reproducible results
  --extra-tests DIR [DIR ...], +i DIR [DIR ...]
                        additonal directories of test images
  --filters {LOF}       additional filters used to put aside generated test
                        inputs that are too far from training data (there is
                        only one filter to choose from for now; the plural is
                        used for future-proofing)
  --norm-factor FLOAT   norm distance upper threshold above which generated
                        inputs are rejected by the oracle (default is 1/4)
  --lb-hard FLOAT       hard lower bound for the distance between original and
                        generated inputs (concolic engine only---default is
                        1/255 for image datasets, 1/100 otherwise)
  --lb-noise FLOAT      extra noise on the lower bound for the distance
                        between original and generated inputs (concolic engine
                        only---default is 1/10)
  --mcdc-cond-ratio FLOAT
                        the condition feature size parameter (0, 1]
  --top-classes CLS     check the top-CLS classifications for models that
                        output estimations for each class (e.g. VGG*)
  --layers LAYER [LAYER ...]
                        test layers given by name or index
  --feature-index INT   to test a particular feature map
  --dbnc-spec SPEC      Feature extraction and discretisation specification
  --dbnc-abstr PKL, --bn-abstr PKL
                        input BN abstraction (.pkl)

The neural network model under tested is specified by --model and a set of raw test data should be given by using --inputs. Some popular datasets like MNIST and CIFAR10 can be directly specified by using the --dataset option directly. --criterion is used to choose the coverage criterion and --norm helps select the norm metric to measure the distance between inputs. Some examples to run DeepConcolic are in the following.

To run an MNIST model

python -m deepconcolic.main --model saved_models/mnist_complicated.h5 --dataset mnist --outputs outs/

To run an CIFAR10 model

python -m deepconcolic.main --model saved_models/cifar10_complicated.h5 --dataset cifar10 --outputs outs/

To test a particular layer

python -m deepconcolic.main --model saved_models/cifar10_complicated.h5 --dataset cifar10 --outputs outs/ --layers 2

To run MC/DC for DNNs on the CIFAR-10 model

python -m deepconcolic.main --model saved_models/cifar10_complicated.h5 --criterion ssc --mcdc-cond-ratio 0.1 --dataset cifar10 --outputs outs

To run Concolic Sign-sign-coverage (MC/DC) for DNNs on the MNIST model

python -m deepconcolic.main --model saved_models/mnist_complicated.h5 --dataset mnist --outputs outs --criterion ssclp

Bayesian Network based Abstraction

To run Concolic BN-based Feature coverage (BFCov) for DNNs on the MNIST model

python -m deepconcolic.main --model saved_models/mnist_complicated.h5 --criterion bfc --norm linf --dataset mnist --outputs outs --dbnc-spec dbnc/example.yaml

See the example YAML specification for details on how to configure the BN-based abstraction.

To run Concolic BN-based Feature-dependence coverage (BFdCov) for DNNs on the MNIST model

python -m deepconcolic.main --model saved_models/mnist_complicated.h5 --criterion bfdc --norm linf --dataset mnist --outputs outs --dbnc-spec dbnc/example.yaml

You could adjust the following two parameters in the DBNC specification file defined by --dbnc-spec to dump the generated bayesian network to files bn4trained.yml and bn4tests.yml.

   dump_bn_with_trained_dataset_distribution: True,
   dump_bn_with_final_dataset_distribution: True,

Fuzzing Engine

DeepConcolic additionally features an experimental fuzzing engine. The following command illustrates how to exercise this engine on a classifier for the CIFAR10 dataset: it will generate at most 1000 images obtained by mutating inputs randomly drawn from the CIFAR10 validation dataset, and save them into the outs/cifar10-fuzzing-basic directory. Aversarial examples can be identified in the latter directory by searching for files named <test-number>-adv-<wrong-label>.png, derived from file <test-number>-original-<true-label>.png. Passed tests are named in a similar way, as <test-number>-ok-<label>.png.

python3 -m deepconcolic.fuzzer --dataset cifar10 --model saved_models/cifar10_complicated.h5 --processes 2 --outputs outs/cifar10-fuzzing-basic -N 1000

Further options are available to use this engine. It is for instance possible to specify a set of files used as seeds for fuzzing with the option --inputs, as in:

python3 -m deepconcolic.fuzzer --dataset mnist --model saved_models/mnist_complicated.h5  --inputs data/mnist-seeds --processes 5 --outputs outs/mnist-fuzzing-given-seeds -N 1000

or sample N inputs from the validation dataset beforehand with --sample N:

python3 -m deepconcolic.fuzzer --dataset cifar10 --model saved_models/cifar10_complicated.h5 --sample 10 --processes 5 --outputs outs/cifar10-fuzzing-sub-sample10 -N 1000

Working with Your Own Datasets

DeepConcolic provides means for working with additional datasets, that can be provided via a dedicate plugin system. Such plugins are Python modules that are loaded when the tool starts, and are searched within any directory listed in the colon-separated environment variable DC_PLUGINS_PATH if this variable is defined, or else within the ./dc_plugins directory if it exists (note the latter is relative to the current working directory).

Then, a new dataset can be registered by calling the deepconcolic.datasets.register_dataset function with a name for the dataset as first argument, and a function that loads and returns a dataset description as second argument. The latter function must accept any set of named arguments (for future extensions), and return a tuple with: (i) a pair of arrays containting trainting data and labets; (ii) a similar pair for validation; (iii) the shape of each individual input element; (iv) a descriptor string in {image, normalized, unknown} (used for determining the input feature encoding—note the format of this descriptor is likely to be refined in future versions); and (v) a list of strings showing the individual label names. The dataset arrays can be given using numpy.ndarray or pandas.Dataframe dataframes. The typical pattern is as follows (for loading, e.g., the MNIST dataset provided by tensorflow, as already done in deepconcolic.datasets):

def load_mnist_data (**_):
  import tensorflow as tf
  img_shape = 28, 28, 1
  (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data ()
  x_train = x_train.reshape (x_train.shape[0], *img_shape).astype ('float32') / 255
  x_test = x_test.reshape (x_test.shape[0], *img_shape).astype ('float32') / 255
  return (x_train, y_train), (x_test, y_test), img_shape, 'image', \
         [ str (i) for i in range (0, 10) ]
register_dataset ('mnist', load_mnist_data)

For further illustrative purposes, we provide an example dataset plugin, which can be used to randomly generate classification tasks. This plugin registers several datasets (named, e.g., rand10_2, rand10_5, and rand100_5) upon startup of DeepConcolic, which should then show as valid options for the --dataset option. We also provide a utility script to construct and train small DNNs for the above toy datasets: To train a classifier for the rand10_2 dataset, and then print a short classification report:

# The following saves the trained model under `/tmp' on Unix-style systems:
python3 -m utils.train4random rand10_2
python3 -m deepconcolic.eval_classifier --dataset rand10_2 --model /tmp/rand10_2_dense_50_50_10_10.h5

To run the fuzzer on the newly trained model, using a sample of 10 initial test data and 5 processes:

python3 -m deepconcolic.fuzzer --dataset rand10_2 --model /tmp/rand10_2_dense_50_50_10_10.h5 --sample 10 --processes 5 --outputs outs/rand10_2-fuzz1000 -N 1000

The above command outputs new inputs within a file outs/rand10_2-fuzz1000/new_inputs.csv.

Tool 2 -- testRNN: Coverage Guided Testing for Recurrent Neural Networks

For long short-term memory models (LSMTs), we design new coverage metrics to consider the internal behaviour of the LSTM layers in processing sequential inputs. We consider not only the tighter metric that quantifies the temporal behaviour (i.e., temporal coverage) but also some looser metrics that quantify either the gate values (i.e., Neuron Coverage and Boundary Coverage) or value change in one step (i.e., Stepwise Coverage).

The paper is available at https://arxiv.org/pdf/1911.01952.pdf.

Four coverage test metrics are applicable:

1. Neuron Coverage (NC),
2. Boundary Coverage (BC),
3. Stepwise Coverage (SC),
4. Temporal Coverage (TC)

A few pre-trained LSTM models:

1. Fashion-MNIST
2. Sentiment Analysis,
3. MNIST Handwritten Digits,
4. UCF101 (need to download and put into the dataset file)

As running example, we download the pre-trained Fasion-MNIST model as follows.

wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/fashion_mnist_lstm.h5

Command to Run:

We have two commands to run testing procedure and to run result analysis procedure, respectively.

to run testing procedure

python -m testRNN.main --model <modelName>
                           --TestCaseNum <Num. of Test Cases>
                           --threshold_SC <SC threshold>
                           --threshold_BC <BC threshold>
                           --symbols_TC <Num. of symbols>
                           --seq <seq in cells to test>
                           --mode <modeName>
                           --outputs <output directory>

where:

1. <modelName> is in {sentiment, mnist, fashion_mnist, ucf101}
2. <Num. of Test Cases> is the expected number of test cases
3. <Mutation Method> is in {random, genetic}
4. <SC threshold> is in [0, 1]
5. <BC threshold> is in [0, 1]
6. <Num. of symbols> is in {1, 2, 3...}
7. <seq in cells to test> is in {mnist: [4, 24], fashion_mnist: [4, 24], sentiment: [400, 499], ucf101: [0, 10]}
8. <modeName> is in {train, test} with default value test
9. <output directory> specifies the path of the directory to save the output record and generated examples

For example, we can run the following

python -m testRNN.main --model fashion_mnist --TestCaseNum 10000 --Mutation random --threshold_SC 0.6 --threshold_BC 0.7 --symbols_TC 3 --seq [4,24] --outputs testRNN_output

which says that, we are working with Fashion-MNIST model, and the genetic algorithm based test case generation will terminate when the number of test cases is over 10000. We need to specify other parameters including threshold_SC, threshold_BC, symbols_TC, and seq. Moreover, the log is generated to the file testRNN_output/record.txt. Also the output of adversarial examples can be found in testRNN_output/adv_output

Tool 3 -- EKiML: Embedding Knolwedge into Tree Ensembles

In this tool, we consider embedding knowledge into machine learning models. The knowledge expression we considered can express e.g., robustness and resilience to backdoor attack, etc. That is, we can "embed" knowledge into a tree ensemble, representing a backdoor attack on the tree ensemble. Also, we can "detect" if a tree ensemble has been attacked.

The paper is available at https://arxiv.org/pdf/2010.08281.pdf.

Download pre-trained models

As the running example, we download the pre-trained HAR tree model as follows.

wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/har_tree_black-box.npy
wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/har_forest_black-box.npy

Command to Run

python -m EKiML.main --Dataset <DatasetName>
		     --Mode <modeName>
		     --Embedding_Method <embeddingMethod>
		     --Model <modeType>
		     --Pruning <pruningFlag>
		     --SaveModel <saveModelFlag>
		     --workdir <workDirectory>

where the flags have multiple options:

1. <DatasetName> is in {'iris', 'har', 'breast_cancer', 'mushroom', 'nursery, 'cod-rna', 'sensorless', 'mnist'}.
2. <modeName> is in {'embedding', 'synthesis'}, where 'synthesis' denotes the "extraction".
3. <embeddingMethod> is in {'black-box', 'white-box'}
4. <modeType> is in {'forest', 'tree'}
5. <pruningFlag> is in {True, False}, with default value False
6. <saveModelFlag> is in {True, False}, with default value False
7. <workDirectory> is the working directory, with default value 'EKiML_workdir'
8. <Datadir> is the directory where dataset files are located (default is 'EKiML/dataset')

For example, we can run the following

python -m EKiML.main --Dataset har --Mode synthesis --Embedding_Method black-box --Model tree --workdir 'EKiML_har' --Datadir 'datasets'

which suggests that we are considering the HAR dataset, tryng to synthesise knowledge from a pre-trained tree by applying our black-box synthesis algorithm.

Tool 4 -- GUAP: Generalised Universal Adversarial Perturbation

Tool for generating spatial-transfermed or additive universarial perturbations, the paper 'Generalizing Universal Adversarial Attacks Beyond Additive Perturbations' was accepted by ICDM 2020.

Please cite Yanghao Zhang, Wenjie Ruan, Fu Wang, and Xiaowei Huang, Generalizing Universal Adversarial Attacks Beyond Additive Perturbations, The IEEE International Conference on Data Mining (ICDM 2020), November 17-20, 2020, Sorrento, Italy

The paper is avaiable at: https://arxiv.org/pdf/2010.07788.pdf

In this paper, for the first time we propose a unified and flexible framework, which can capture the distribution of the unknown additive and non-additive adversarial perturbations jointly for crafting Generalized Universal Adversarial Perturbations. Specifically, GUAP can generate either additive (i.e., l_inf-bounded) or non-additive (i.e., spatial transformation) perturbations, or a combination of both, which considerably generalizes the attacking capability of current universal attack methods.

Colab demo:

There is also a notebook demo Colab_GUAP.ipynb, which can be run on the Colab.

Running environment:

pip install torch torchvision matplotlib

Download target Models

wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/cifar10_vgg19.pth
wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/cifar10_resnet101.pth
wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/cifar10_dense121.pth
wget -P saved_models https://cgi.csc.liv.ac.uk/~acps/models/fashion_mnist_modela.pth

Command to Run

(from within the GUAP sub-directory)

usage: run_xxxxxx.py [-h] [--dataset DATASET] [--lr LR]
                            [--batch-size BATCH_SIZE] [--epochs EPOCHS]
                            [--l2reg L2REG] [--beta1 BETA1] [--tau TAU]
                            [--eps EPS] [--model MODEL]
                            [--manualSeed MANUALSEED] [--gpuid GPUID] [--cuda]
                            [--resume] [--outdir OUTDIR]

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Fashion-MNIST
  --lr LR               Learning rate
  --batch-size BATCH_SIZE
  --epochs EPOCHS       number of epochs to train for
  --l2reg L2REG         weight factor for l2 regularization
  --beta1 BETA1         beta1 for adam
  --tau TAU             max flow magnitude
  --eps EPS             allow for linf noise
  --model MODEL         modelA
  --manualSeed MANUALSEED
                        manual seed
  --gpuid GPUID         multi gpuid
  --cuda                enables cuda
  --resume              load pretrained model
  --outdir OUTDIR       output dir

Generalizing UAP for Fashion_MNIST:

python run_fashion_mnist.py --cuda --gpuid 0 --resume

Generalizing UAP for Cifar10:

python run_cifar.py --cuda --gpuid 0 --model VGG19 --tau 0.1 --eps 0.03

Generalizing UAP for ImageNet:

python run_imagenet.py --cuda --gpuid 0,1 --model ResNet152 --tau 0.1 --eps 0.03

Experimental results:

Publications

For DeepConcolic,

@inproceedings{swrhkk2018,
  AUTHOR    = { Sun, Youcheng
                and Wu, Min
                and Ruan, Wenjie
                and Huang, Xiaowei
                and Kwiatkowska, Marta
                and Kroening, Daniel },
  TITLE     = { Concolic Testing for Deep Neural Networks },
  BOOKTITLE = { Automated Software Engineering (ASE) },
  PUBLISHER = { ACM },
  PAGES     = { 109--119 },
  ISBN      = { 978-1-4503-5937-5 },
  YEAR      = { 2018 }
}

@article{sun2018testing,
  AUTHOR    = { Sun, Youcheng
                and Huang, Xiaowei
                and Kroening, Daniel },
  TITLE     = { Testing Deep Neural Networks },
  JOURNAL   = { arXiv preprint arXiv:1803.04792 },
  YEAR      = { 2018 }
}

@article{10.1145/3358233,
author = {Sun, Youcheng and Huang, Xiaowei and Kroening, Daniel and Sharp, James and Hill, Matthew and Ashmore, Rob},
title = {Structural Test Coverage Criteria for Deep Neural Networks},
year = {2019},
issue_date = {October 2019},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {18},
number = {5s},
issn = {1539-9087},
url = {https://doi.org/10.1145/3358233},
doi = {10.1145/3358233},
journal = {ACM Trans. Embed. Comput. Syst.},
articleno = {Article 94},
numpages = {23},
keywords = {test criteria, Neural networks, test case generation} }

For testRNN,

@article{testRNN,
  author    = {Wei Huang and
               Youcheng Sun and
	       Xingyu Zhao and 
               James Sharp and
	       Wenjie Ruan and 
	       Jie Meng and 
               Xiaowei Huang},
  title     = {Coverage Guided Testing for Recurrent Neural Networks},
  journal   = {IEEE transactions on Reliability},
  year      = {2021},
}

For EKiML,

@article{DBLP:journals/corr/abs-2010-08281,
  author    = {Wei Huang and
               Xingyu Zhao and
               Xiaowei Huang},
  title     = {Embedding and Synthesis of Knowledge in Tree Ensemble Classifiers},
  journal   = {Machine Learning},
  year      = {2021}
}

For GUAP,

@inproceedings{zhang2020generalizing,
      title={Generalizing Universal Adversarial Attacks Beyond Additive Perturbations},
      author={Yanghao Zhang and Wenjie Ruan and Fu Wang and Xiaowei Huang},
      year={2020},
      booktitle = {ICDM 2020}
}

deepconcolic's Issues

issues registering custom dataset

We have tried to upload our own dataset but were faced with some problems like register_dataset not being recognized by our system. Is there a clear way to easily upload a custom dataset. Thank you so much

Is batch normalization layer supported?

Hello,

Thank you for the awesome package. I found it seems that the library does not support a batch normalization layer in a model. Since the batch normalization is quite useful, is there any future plans for supporting it?

issues on running examples in ASE-experiments

Hello, sorry to bother, I encountered some problems while running a sample command in the ASE experiments folder，
python MNIST_nc.py ../../saved_models/mnist_complicated.h5
Is there no such file named mnist_complicated.h5 in this folder, or is there something wrong with my operation? How can I get this file
If this file does exist,？can I run it with models in other folders as parameters？

作者您好，打扰您了，我在运行ASE-experiments文件夹下的命令的时候遇到了一些问题
python MNIST_nc.py ../../saved_models/mnist_complicated.h5
提示文件不存在，这个文件是本来就没有的，还是因为我的操作问题，
导致没有下载到这个文件，我该如何获得这个文件
如果这个文件不存在，可不可以用别的h5文件代替

'test_objectt' object has no attribute 'eval_batch'

Thank you for providing this open source attack for the community. I tried to replicate your attack but was unable to do so, could you help me troubleshoot this?

I tried running the MNIST model as per the instructions in the readme.
python deepconcolic.py --model ../saved_models/mnist_complicated.h5 --mnist-data --outputs outs/`

But unfortunately, I had this error.

Do you know how I can mitigate this error? Would really appreciate your help in this!

Cannot download models, model_fashion_mnist.h5

Hi, this is great work, thanks for your kind share.

But I cannot download the models. Would you please share them?

Thanks.

How to obtain the coverage of the test input set T without generating adversarial samples?

Hello~
Could I use deepconcolic to obtain the coverage of the test input set T without generating adversarial samples? I just want to get the coverage of a test set T under the sign-sign criterion.

Thanks in advance!

Unable to use Deep concolic with custom cnn model

Hi,
I have trained a simple CNN and trying to generate test data using Deep concolic but when in run deepconcolic.py on my cnn model, all the layers are recognized as not testable.
Following is abstract form command line window:

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 90, 110, 32)       896       
_________________________________________________________________
activation (Activation)      (None, 90, 110, 32)       0         
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 45, 55, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 43, 53, 32)        9248      
_________________________________________________________________
activation_1 (Activation)    (None, 43, 53, 32)        0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 21, 26, 32)        0         
_________________________________________________________________
flatten (Flatten)            (None, 17472)             0         
_________________________________________________________________
dense (Dense)                (None, 64)                1118272   
_________________________________________________________________
activation_2 (Activation)    (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 40)                2600      
_________________________________________________________________
activation_3 (Activation)    (None, 40)                0         
=================================================================
Total params: 1,131,016
Trainable params: 1,131,016
Non-trainable params: 0
_________________________________________________________________
Loading input data from data/attdata... data/attdata
data/attdata

...

400 loaded.
DNN under test has 11 layer functions, 0 of which is to be covered:
[]

Starting tests for criterion NC with norm L0.                                   
Outputs will go into: attout/
Reporting into: attout/NC_L0-2020-12-16-00-04-03.445382_report.txt              
Randomly selecting an input from test data.                                     
#0 NC: 0.00000000%   

Please suggest how to resolve this issue.

Thanks,
Lalit Kumar

Unable to download the model in example

Hi,
I trying to use the tool, but I cannot download the model in the example.
Thank you

About the running time of ssclp

Hello~ When I ran this command (python deepconcolic.py —model ../saved_models/mnist_complicated.h5 —criterion ssclp —mnist-data —outputs outs) to run Concolic Sign-sign-coverage (MC/DC) for DNNs on the MNIST model, I ran it for two days and still stuck in the place where the constraints were solved, as shown in the figure below.

The sequential we got doesn't match our CNN sequential

This is our CNN's "sequential"

And this is what we got by running this command "python -m deepconcolic.main --outputs outs/ReddRefit/ssclp-lo --dataset refitRedd --model models/cnn.h5 --criterion ssclp --norm l0 --save-all-tests --max-iterations 100"

The LP model is always NOT feasible when using Python3

Neuron Boundary Coverage not supported

Hello, the criterion argument has 'bc' in README as well as in the paper, but in implementation, the NBC is not supported.

Why use a BUFFER_SIZE constant in the function update_nc_map_via_inst?

Hello, I have some questions when review the update_nc_map_via_inst function in utils.py.
I'm not quite sure the purpose of constant BUFFER_SIZE .

In my opinion, each element in clayers[i].activations represents an activation pattern produced by an instance. So when we update nc_map via current instance, why should we have operations on current nc_map and previous activation pattern?

Thanks a lot for your kindly assistance.

def update_nc_map_via_inst(clayers, activations, layer_feature=None):
  for i in range(0, len(clayers)):
    ## to get the act of layer 'i'
    act=copy.copy(activations[clayers[i].layer_index])
    ## TODO
    if act_in_the_layer(clayers[i].layer)=='relu':
      act[act==0]=MIN/10
    act[act>=0]=0
    if clayers[i].nc_map is None: ## not initialized yet
      clayers[i].initialize_nc_map(layer_feature)
      clayers[i].nc_map=np.logical_and(clayers[i].nc_map, act)
    else:
      clayers[i].nc_map=np.logical_and(clayers[i].nc_map, act)
    ## update activations after nc_map change
    if len(clayers[i].activations)>=BUFFER_SIZE:
      ind=np.random.randint(0,BUFFER_SIZE)
      clayers[i].activations[ind]=(act)
    else:
      clayers[i].activations.append(act)
    ## clayers[i].update_activations()
    for j in range(0, len(clayers[i].activations)):
      clayers[i].activations[j]=np.multiply(clayers[i].activations[j], clayers[i].nc_map)
      clayers[i].activations[j][clayers[i].activations[j]>=0]=MIN