Official code for "Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems" in CVPR 2025.

Requirement:

pip install -r requirements.txt

Dataset:

For convenience, we put the dataset utilized in our experiment in this link. Please put the four folders in the root folder.

1. CIFAR10
2. CIFAR100
3. Facescrub
4. TinyImagenet

Trained Model

Download our trained models: please download our trained models via this link

Conduct the experiments

1. Train the classifier and run decoing-based MIA attack (for inference only, please comment the training part) bash run_exp.sh
2. Run the GAN-based MIA attack python run_gan_attack.py

Important hyperparameters in run_exp.sh file

1. arch: vgg11_bn_sgm (this for CIFAR10, 100, and facescrub, bn means batchnorm and sgm means Sigmoid activation) and resnet20 (for Tinyimagenet)
2. cutlayer_list: which layer to split the encoder and decoder;
3. AT_regularization: The defense method for split learning; gan_adv for adversarial training, dropout for dropout defense, topkprune for topkpruning, pruning for PATROL. For example, gan_adv_step1_pruning180 means using adversarial training with PATROL pruning at 180 epochs.
4. AT_regularization_strength: The weight for the defense method
5. dataset_list: dataset for experiment: cifar10, cifar100, facescrub and Tinyimagenet
6. regularization: Gaussian means adding the noise corruption mentioned in CEM
7. regularization_strength_list: the variance of the noise corruption, can be a list like: "0.01 0.025 0.05 0.1 0.15"
8. lambd_list: weight for CEM regularization strength (default 8 or 16); 0 without using CEM
9. log_entropy: for log optimization in the loss function.
10. bottleneck_option_list means adding the bottleneck layer after the encoder

Cite the work:

@inproceedings{xia2025theoretical,
  title={Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems},
  author={Xia, Song and Yu, Yi and Yang, Wenhan and Ding, Meiwen and Chen, Zhuo and Duan, Ling-Yu and Kot, Alex C and Jiang, Xudong},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={8753--8763},
  year={2025}
}