Compressing Adversarial Attacks

Code for paper : https://arxiv.org/abs/2209.13785 published at FICC '23 and Springer

Downloading Models:

Please run download_checkpoint.sh to download the models and use environment.yml to setup the environment using conda

Download models from drive link https://drive.google.com/drive/folders/1ZJNrhe6IW82tl4VD0X84Scs4ocETL8-1?usp=sharing and store them in /dvit/pretrained and checkpoints respectively

Running Experiments

final_runner.ipynb: is all you need to run all the experiments Description of its commands:

!bash ./experiments/eval_models.sh - evaluate all the models using eval_models.py

!bash ./experiments/attack_uap_swap.sh - create attacks on base and distilled models using attack_uap.py

!bash ./experiments/attack_uap_pruning.sh - create attacks on pruned models using attack_uap.py

!bash ./experiments/attack_boundary_swap.sh - create black box attacks on quantized using attack_fb_swap.py

utils.py - loads all the models