Indirect Local Attacks for Context-aware Semantic Segmentation Networks

Introduction

This repository is a PyTorch implementation of indirect local attacks for semantic segmentation paper(https://arxiv.org/abs/1911.13038) accepted at ECCV 2020. The code is easy to use for attacking various datasets. The codebase mainly uses ResNet50 as backbone and can be easily adapted to other basic classification structures.

Usage

1. Download the Cityscapes sampled demo images and pretrained public models

   cd pretrained/
   bash download_dataset.sh
   bash download_pretrained_models.sh
   

TO BE ADDED SOON

2. Adaptive attack a sample image on a pretrained model with a given sparsity level. We have evaluated attacks on 4 architecures - fcn50, psanet50, pspnet50, danet50.

   sh scripts/adaptive_attack.sh cityscapes psanet50 1.jpg 0.75

Citation

If you find the code useful, please consider citing:

@misc{indirectattacks2020,
  author={Krishna Kanth Nakka, Mathieu Salzmann},
  title={indirectattacks},
  howpublished={\url{https://arxiv.org/abs/1911.13038}},
  year={2020}
}