#Disguised-Nets#

Deep learning over perturbed data

A. Perturb training and testing csv files with:
 1) Block-wise RMT 
 2) Permutation 
 3) Random Additive Perturbation
B. Run the MNIST CNN over different block sizes and noise levels

Dependencies

numpy
pandas
TensorFlow

Datasets

download ubyte.gz files from http://yann.lecun.com/exdb/mnist

Run: python read_mnist.py from ./datasets/

Individual Program Flows

A. Perturbation flow --- > Disguising training and testing images
	Run: python ./Perturb/main.py data_path/data_name image_size block_size noise_level RMT_type
	-
	Example: python ./Perturb/main.py ./dataset/mnist 28 28 4 4 10.0 orthogonal
	-
	Outputs: data_name_{train/test}_${noise}_${block_dim}_${block_dim}.csv 

** there is no padding so ensure the image size is a multiple of block_size
B. DNN over Disguised Images
	Run: python ./models/conv_simple.py training_csv testing_csv
	-
	Example: python ./models/conv_simple.py fashion_train.csv fashion_test.csv
	Example: python ./models/conv_simple.py train_10.0_4.csv test_10.0_4.csv

	Generates a file called submission.csv
	
Run: python ./models/test_results.py

Produces the test results corect/total

Model Quality and Visual Re-identification Experiments

Run: bash ./run_all_deep.sh

For different block sizes:
    For different noise levels:
        Perturbation flow for MNIST
        for 5 times:
          model quality
          visual re-identification attack

The results for model quality are sequntially written on ./out/org_on_org.csv The results for visual re-identification attack sequentially written on ./out/org_on_pert.csv

Class-membership Attack Experiments

A. Generate class_wise test data (the target class images)

Run: python ./class_membership/class_wise_test.py ./datasets/{mnist/fashion}_test.csv \
  ./datasets/{mnist/fashion}_test_labels.csv

B. Run the class-membershp attacks

1. On untransformed MNIST model

Run: bash ./run_model_dist.sh

2. On transformed MNIST model

Run: bash ./run_model_dist_pert.sh
  
Outputs: The model predictions for each class is written to
./datasets/class_outputs/class_outputs_i for class i

These distributions can be analyzed for entropy/fano factor computation.