"A Statistical Approach to Assessing Neural Network Robustness"

Here you can find a PyTorch implementation of Adaptive Multilevel splitting and code to reproduce the experiments in our paper, "A Statistical Approach to Assessing Neural Network Robustness", to appear at the 7th International Conference on Representation Learning (ICLR 2019).

Instructions

For all experiments:

1. Install the Python package seaborn, and the latest versions of NumPy/PyTorch.
2. Install Gurobi. (This is required for the PLNN package to load, but is not actually used by the core code itself.)
3. Activate Gurobi license.

Experiment 6.2 Ablation Study:

From the base directory:

• python -m exp_6_2.process_data
• python -m exp_6_2.run_exp
• python -m exp_6_2.plot_exp

Experiment 6.3 MNIST/CIFAR10:

From the base directory:

• python -m exp_6_3_mnist.train
• python -m exp_6_3_mnist.run_baseline
• python -m exp_6_3_mnist.run_exp
• python -m exp_6_3_mnist.plot_exp

Repeat for CIFAR10 by replacing mnist with cifar10.

Experiment 6.3 CIFAR100:

This experiment uses a pretrained DenseNet contained in the repo.

From the base directory:

• python -m exp_6_3_cifar100.run_baseline
• python -m exp_6_3_cifar100.run_exp
• python -m exp_6_3_cifar100.plot_exp

Experiment 6.4 Robust Training:

From the base directory:

1. Train the network by the standard method:
   • python -m exp_6_4.train --epochs 100 --starting_epsilon 0.01 --prefix mnist --schedule_length 50 --method 'baseline' --cuda_ids '0'
2. Continue training by the robust method:
   • python -m exp_6_4.train --epochs 100 --starting_epsilon 0.01 --prefix 'mnist_robustified' --schedule_length 50 --method 'robust' --cuda_ids '0' --model_snapshot './snapshots/mnist_baseline_batch_size_50_epochs_100_lr_0.001_opt_adam_real_time_False_seed_0_checkpoint_99.pth'
3. Run and plot the experiment:
   • python -m exp_6_4.run_exp
   • python -m exp_6_4.plot_exp

Change --cuda_ids to your desired GPU card index.

Pertinent files

• /exp_6_2
  • plot_exp.py Produces Figures 1 and 4 from the paper for Experiment 6.2.
  • process_data.py Converts the CollisionDetection property files into pickled objects and a summary file. Also produces naive MC estimates.
  • run_exp.py Runs AMLS on the 500 properties to produce results for Experiment 6.1 and 6.2.
• /exp_6_3_mnist
  • plot_exp.py Produces the MNIST panel of Figure 2.
  • run_baseline.py Produces naive MC estimates of adversarial properties of varying rareness.
  • run_exp.py Runs AMLS on adversarial properties varying the hyperparameters to produce results for Ex 6.3.
  • train.py Trains the simple feedforward classifier on MNIST.
• /exp_6_4
  • plot_exp.py Produces Figures 3 and 5.
  • run_exp.py Runs AMLS on the trained snapshots.
  • train.py Trains the network using standard or robust training, saving snapshots.
• /plnn A modified version of the PLNN package with some bug fixes for the latest PyTorch.
• /convex_adversarial A modified version of the convex_adversarial package.

Note that exp_6_4 contain modified code from here

The directories /exp_6_3_cifar10 and /exp_6_3_cifar100 are analogous to exp_6_3_mnist.

Notes

• Please contact me if you would like to obtain the code for Section 6.1.
• The code to run the experiments was not completely rerun due to the need for computational resources, so contact me if you are having trouble with them.
• You may wish to modify the code so that it runs different combinations of hyperparamaters in parallel across multiple GPUs to reduce the total experiment time.