Repo Overview

This repo contains experimental code for the paper "Recovery Guarantees for Compressible Signals with Adversarial Noise".

Code Overview

• /models: Core codebase for network setup, training, one pixel attack, and other utilities.
• /cleverhans: Forked and modified brach of the Cleverhans library used in conducting the CW, JSMA, and DF attacks.
• /notebooks: Jupyter notebooks for each experiment.

Requirements

• python 2.7.12
• keras 2.2.4, with GPU support
• tensorflow 1.8.0, with GPU support
• numpy(1.16.14), scipy(1.2.2), bottleneck(1.2.1), cvxpy(1.0.24), PIL(5.1.0), jupyter(4.4.0)

Usage

We have created a notebook for every experiment in the paper. In order to reproduce or modify experiments, follow the relevant notebook.