Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning

Reference implementation of the reliable conformal prediction sets (RPS) proposed in the paper:

Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning
Yan Scholten, Stephan Günnemann
International Conference on Learning Representations, ICLR 2025 (Spotlight)
[ Project page | PDF ]

Demo Notebook Example

We provide a demo notebook with an example of how to compute reliable conformal prediction sets (RPS) under calibration poisoning, and a full demo notebook to demonstrate RPS under training and calibration poisoning.

Install

Instructions for dependencies and configurations before running code:

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install -e .
export CUBLAS_WORKSPACE_CONFIG=:4096:8

The code was tested with Python 3.11.9, pip 24.0, PyTorch 2.3.1+cu118, and CUDA 11.8 on NVIDIA GeForce GTX 1080 Ti.

Reproducibility

To start reproducing the results of our paper systematically on a compute cluster, please setup SEML and make sure that the data folder has the following structure:

data/
├── certificates/
├── datasets/
└── models/

1. Training:

Execute the following to train models (which will be stored in data/models/):

seml rcp_training add configs/training/0-ResNet18-CIFAR10.yaml 
seml rcp_training start

2. Calibration and certification:

After model training has completed, compute the reliable prediction sets and certificates (which will be stored in data/certificates/):

seml rcp add configs/rcp/0-cert-setting0.yaml
seml rcp start

To reproduce the full results of our paper, first train the required models using the provided training configurations and then compute the reliable prediction sets and certificates using the provided RCP configurations.

Cite

Please cite our paper if you use this code in your own work:

@inproceedings{scholten2025provably,
    title={Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning},
    author={Yan Scholten and Stephan G{\"u}nnemann},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=ofuLWn8DFZ}
}

Contact

For questions and feedback please contact:

Yan Scholten, Technical University of Munich
Stephan Günnemann, Technical University of Munich

License

The code by Yan Scholten and Stephan Günnemann is licensed under MIT license.