RelaxLoss: Defending Membership Inference Attacks without Losing Utility

Authors: Dingfan Chen, Ning Yu, Mario Fritz

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive evaluations on five datasets with diverse modalities (images, medical data, transaction records), our approach consistently outperforms state-of-the-art defense mechanisms in terms of resilience against MIAs as well as model utility.
Researcher Affiliation Collaboration 1CISPA Helmholtz Center for Information Security 2Salesforce Research 3University of Maryland 4Max Planck Institute for Informatics
Pseudocode Yes Algorithm 1: Relax Loss
Open Source Code Yes Source code is available at https://github.com/DingfanChen/RelaxLoss.
Open Datasets Yes We set up seven target models, trained on five datasets (CIFAR-10, CIFAR-100, CHMNIST, Texas100, Purchase100) with diverse modalities... followed by citations like CIFAR-10 (Krizhevsky et al., 2009). Also mentions We use the preprocessed data provided by Shokri et al. (2017); Song & Mittal (2020)9. and footnote 9 provides https://github.com/inspire-group/membership-inference-evaluation.
Dataset Splits Yes We evenly split each dataset into five folds and use each fold as the training/testing set for the target/shadow model3, and use the last fold for training the surrogate attack model (for Jia et al. (2019); Shokri et al. (2017)).
Hardware Specification Yes Our experiments are conducted with Nvidia Tesla V100 and Quadro RTX8000 GPUs.
Software Dependencies No The paper states 'All our models and methods are implemented in Py Torch' but does not provide specific version numbers for PyTorch or any other software libraries or dependencies.
Experiment Setup Yes We apply SGD optimizer with momentum=0.9 and weight-decay=1e-4 by default. We set the initial learning rate τ = 0.1 and drop the learning rate by a factor of 10 at each decay epoch 11. We list below the decay epochs in square brackets and the total number of training epochs are marked in parentheses: CIFAR-10 and CIFAR-100 [150,225] (300); CH-MNIST [40,60] (80); Texas100 and Purchase100 [50,100] (120).