This project is to integrate the work of Overcoming the Convex Barrier for Simplex Inputs to CROWN algorithm. It cleverly uses the simplex to propose a tighter boundary for the l1 perturbation of the convex activation function network, improving the effect of the CROWN algorithm.
Use compute_bounds interface in SimplexSolver class under simplexcrown/simplex_solver.py to compute bounds. We also provided an example in simple_eval.py
We trained multiple MLP models with different layer using MNIST dataset by mlp_train.py script, the pre-trained models are under /models directory. You can use mnist.py script to perform experiments on mnist.
This work is based on following papers:
@article{behl2021overcoming,
title={Overcoming the convex barrier for simplex inputs},
author={Behl, Harkirat Singh and Kumar, M Pawan and Torr, Philip and Dvijotham, Krishnamurthy},
journal={Advances in Neural Information Processing Systems},
volume={34},
pages={4871--4882},
year={2021}
}
@article{zhang2018efficient,
title={Efficient Neural Network Robustness Certification with General Activation Functions},
author={Zhang, Huan and Weng, Tsui-Wei and Chen, Pin-Yu and Hsieh, Cho-Jui and Daniel, Luca},
journal={Advances in Neural Information Processing Systems},
volume={31},
pages={4939--4948},
year={2018},
url={https://arxiv.org/pdf/1811.00866.pdf}
}
@article{xu2020automatic,
title={Automatic perturbation analysis for scalable certified robustness and beyond},
author={Xu, Kaidi and Shi, Zhouxing and Zhang, Huan and Wang, Yihan and Chang, Kai-Wei and Huang, Minlie and Kailkhura, Bhavya and Lin, Xue and Hsieh, Cho-Jui},
journal={Advances in Neural Information Processing Systems},
volume={33},
year={2020}
}
Please notice that simplexcrown/lirpa directory is the code for original method provided in Overcoming the Convex Barrier for Simplex Inputs. Also simplexcrown/crown.py, simplexcrown/linear.py, simplexcrown/relu.py are based on the code [email protected]:huanzhang12/ECE584-SP24-assignment2.git from @huanzhang12 and @schawla7.