This repository contains a reference implementation for our defense against adversarial examples, which uses a technique from conformal prediction called inductive Venn-ABERS predictors (IVAPs; see [1, 2, 3] for more details). This work has been published in the proceedings of the ESANN 2019 conference [4].
In order to run our code, it is necessary to install the following dependencies:
- Python 3.5.2 or higher (https://www.python.org/);
- TensorFlow 1.8.0 (https://www.tensorflow.org/);
- Keras 2.2.0 (https://www.keras.io/);
- NumPy 1.14.5 (http://www.numpy.org/);
- Foolbox 1.4.0 (https://foolbox.readthedocs.io).
There is a Jupyter Notebook cifar10_demo.ipynb in this repository which contains detailed explanations and demonstrations of how to use the IVAP defense on a subset of the CIFAR-10 data set.
- Vovk, Vladimir, Ivan Petej, and Valentina Fedorova. "Large-scale probabilistic predictors with and without guarantees of validity." Advances in Neural Information Processing Systems. 2015. PDF
- Vovk, Vladimir, Alex Gammerman, and Glenn Shafer. Algorithmic learning in a random world. Springer Science & Business Media, 2005.
- Shafer, Glenn, and Vladimir Vovk. "A tutorial on conformal prediction." Journal of Machine Learning Research 9.Mar (2008): 371-421. PDF
- Peck, Jonathan, Bart Goossens and Yvan Saeys. "Detecting adversarial examples with inductive Venn-ABERS predictors." European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning. 2019. PDF
