Greedy Adversarial Pruning

Official repo for the paper

Hardening DNNs against Transfer Attacks during Network Compression using Greedy Adversarial Pruning

Authors: Jonah O'Brien Weiss, Tiago Alves, and Sandip Kundu

Published in the 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS)

---

Getting Started

Prerequisites

• Conda
• Python 3.9

Installation

Run inshall.sh. This will create a conda environment called "gap", install the requirements, and setup the submodule dependency on the "shrinkbench" repo.

Running Experiments

1. Modify src/config.yaml to define what experiments to run. This is prepopulated with an example configuration in debug mode. Configs get saved in a new file when running experiements.
2. Run download_dataset.py
3. Run run.py.

---

Folder Structure

Notable files under src:

src/
├── analyze_data.py           # Scripts for plotting experiment results and data
├── config.yaml               # Main configuration file for experiments
├── download_dataset.py       # Script to download required datasets
├── email_pw.txt              # (Sensitive) Email password for notifications (should not be shared)
├── evaulate_model.py         # Model evaluation utilities and scripts
├── experiments.py            # Main experiment orchestration logic
├── remove_models.py          # Script to remove or clean up models
├── run.py                    # Entry point to run experiments
└── shrinkbench/              # Submodule: ShrinkBench library for pruning and compression
    ├── experiment/          # Core experiment tracking and logging functionality
    ├── models/              # Neural network model definitions and utilities
    ├── pruning/             # Base pruning classes and adversarial pruning implementations
    └── strategies/          # Different pruning strategy implementations (magnitude, gradient, etc.)

---

Contact

For questions or issues, please contact jgow98@gmail.com.