NICGSlowDown is designed to generate efficiency adversarial examples to evaluate the efficiency robustness of NICG models. The generated adversarial examples are realistic and human-unnoticable images while consume more computational resources than benign images.
Our approach overview is shown in the above figure, for the detail design, please refer to our papers.
- src -main source codes.
- ./src/model -the model architecture of the NICG models.
- ./src/attack -the implementation of proposd attack algorithm.
- train.py -the script to train the NICG models.
- generate_adv.py -this script generate the adversarial examples.
- test_latency.py -this script measure the latency of the generated adversarial examples.
- gpu4.sh -bash script to generate adversarial examples and measure the efficiency.
We provide the bash script that generate adversarial examples and measure the efficiency in gpu4.sh. gpu5.sh, gpu6.sh,gpu7.sh are implementing the similar functionality but for different gpus.
So just run bash gpu4.sh
The above figure shows the efficiency distribution of the benign images and the adversarial images. The ares under the cumulative distribution function (CDF) represent the victim NICG models' efficiency. A large area implies the model is less efficiency.
The first row shows the benign images and the second row shows the generated efficiency adversarial images.
If you find this repository is helpful to you, please consider cite
@inproceedings{chen2022nicgslowdown,
title={NICGSlowDown: Evaluating the Efficiency Robustness of Neural Image Caption Generation Models},
author={Chen, Simin and Song, Zihe and Haque, Mirazul and Liu, Cong and Yang, Wei},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={15365--15374},
year={2022}
}