Recently, there has been much progress on adversarial attacks against neural networks, such as the cleverhans library and the code by Carlini and Wagner. We now complement these advances by proposing an attack challenge. We have trained a robust network, and the objective is to find a set of adversarial examples on which this network achieves only a low accuracy. To train an adversarially-robust network, we followed the approach from our recent paper:
Towards Deep Learning Models Resistant to Adversarial Attacks
Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu
https://arxiv.org/abs/1706.06083.
As part of the challenge, we release both the training code and the network architecture, but keep the network weights secret. We invite any researcher to submit attacks against our model (see the detailed instructions below). We will maintain a leaderboard of the best attacks for the next two months and then publish our secret network weights.
The goal of our challenge is to clarify the state-of-the-art for adversarial robustness on MNIST. Moreover, we hope that future work on defense mechanisms will adopt a similar challenge format in order to improve reproducibility and empirical comparisons.
| Attack | Submitted by | Accuracy | Submission Date |
|---|---|---|---|
| PGD against three independently and adversarially trained copies of the network |
Florian Tramèr | 93.54% | Jul 5, 2017 |
| FGSM on the CW loss for model B from "Ensemble Adversarial Training [...]" |
Florian Tramèr | 94.36% | Jun 29, 2017 |
| FGSM on the CW loss for the naturally trained public network |
(initial entry) | 96.08% | Jun 28, 2017 |
| PGD on the cross-entropy loss for the naturally trained public network |
(initial entry) | 96.81% | Jun 28, 2017 |
| FGSM on the cross-entropy loss for the adversarially trained public network |
(initial entry) | 97.66% | Jun 28, 2017 |
| PGD on the cross-entropy loss for the adversarially trained public network |
(initial entry) | 97.79% | Jun 28, 2017 |
The objective of the challenge is to find black-box (transfer) attacks that are effective against our MNIST model.
Attacks are allowed to perturb each pixel of the input image by at most epsilon=0.3.
To ensure that the attacks are indeed black-box, we release our training code and model architecture, but keep the actual network weights secret.
We invite any interested researchers to submit attacks against our model. The most successful attacks will be listed in the leaderboard above. As a reference point, we have seeded the leaderboard with the results of some standard attacks.
We used the code published in this repository to produce an adversarially robust model for MNIST classification. The model is a convolutional neural network consisting of two convolutional layers (each followed by max-pooling) and a fully connected layer. This architecture is derived from the MNIST tensorflow tutorial.
The network was trained against an iterative adversary that is allowed to perturb each pixel by at most epsilon=0.3.
The random seed used for training and the trained network weights will be kept secret.
The sha256() digest of our model file is:
14eea09c72092db5c2eb5e34cd105974f42569281d2f34826316e356d057f96d
We will release the corresponding model file on September 1st 2017, which is roughly two months after the start of this competition.
We are interested in adversarial inputs that are derived from the MNIST test set.
Each pixel can be perturbed by at most epsilon=0.3 from its initial value.
All pixels can be perturbed independently, so this is an l_infinity attack.
Each attack should consist of a perturbed version of the MNIST test set. Each perturbed image in this test set should follow the above attack model.
The adversarial test set should be formated as a numpy array with one row per example and each row containing a flattened
array of 28x28 pixels.
Hence the overall dimensions are 10,000 rows and 784 columns.
Each pixel must be in the [0,1] range.
See the script pgd_attack.py for an attack that generates an adversarial test set in this format.
In order to submit your attack, save the matrix containing your adversarial examples with numpy.save and email the resulting file to mnist.challenge@gmail.com.
We will then run the run_attack.py script on your file to verify that the attack is valid and to evaluate the accuracy of our secret model on your examples.
After that, we will reply with the predictions of our model on each of your examples and the overall accuracy of our model on your evaluation set.
If the attack is valid and outperforms all current attacks in the leaderboard, it will appear at the top of the leaderboard. Novel types of attacks might be included in the leaderboard even if they do not perform best.
We strongly encourage you to disclose your attack method. We would be happy to add a link to your code in our leaderboard.
The code consists of six Python scripts and the file config.json that contains various parameter settings.
python train.py: trains the network, storing checkpoints along the way.python eval.py: an infinite evaluation loop, processing each new checkpoint as it is created while logging summaries. It is intended to be run in parallel with thetrain.pyscript.python pgd_attack.py: applies the attack to the MNIST eval set and stores the resulting adversarial eval set in a.npyfile. This file is in a valid attack format for our challenge.python run_attack.py: evaluates the model on the examples in the.npyfile specified in config, while ensuring that the adversarial examples are indeed a valid attack. The script also saves the network predictions inpred.npy.python fetch_model.py name: downloads the pre-trained model with the specified name (at the momentadv_trainedornatural), prints the sha256 hash, and places it in the models directory.
Model configuration:
model_dir: contains the path to the directory of the currently trained/evaluated model.
Training configuration:
random_seed: the seed for the RNG used to initialize the network weights.max_num_training_steps: the number of training steps.num_output_steps: the number of training steps between printing progress in standard output.num_summary_steps: the number of training steps between storing tensorboard summaries.num_checkpoint_steps: the number of training steps between storing model checkpoints.training_batch_size: the size of the training batch.
Evaluation configuration:
num_eval_examples: the number of MNIST examples to evaluate the model on.eval_batch_size: the size of the evaluation batches.eval_on_cpu: forces theeval.pyscript to run on the CPU so it does not compete withtrain.pyfor GPU resources.
Adversarial examples configuration:
epsilon: the maximum allowed perturbation per pixel.k: the number of PGD iterations used by the adversary.a: the size of the PGD adversary steps.random_start: specifies whether the adversary will start iterating from the natural example or a random perturbation of it.loss_func: the loss function used to run pgd on.xentcorresponds to the standard cross-entropy loss,cwcorresponds to the loss function of Carlini and Wagner.store_adv_path: the file in which adversarial examples are stored. Relevant for thepgd_attack.pyandrun_attack.pyscripts.
After cloning the repository you can either train a new network or evaluate/attack one of our pre-trained networks.
- Start training by running:
python train.py
- (Optional) Evaluation summaries can be logged by simultaneously running:
python eval.py
- For an adversarially trained network, run
python fetch_model.py adv_trained
and use the config.json file to set "model_dir": "models/adv_trained".
- For a naturally trained network, run
python fetch_model.py natural
and use the config.json file to set "model_dir": "models/natural".
- Create an attack file by running
python pgd_attack.py
- Evaluate the network with
python run_attack.py