MExMI: Pool-based Active Model Extraction Crossover Membership Inference

Authors: Yaxin Xiao, Qingqing Ye, Haibo Hu, Huadi Zheng, Chengfang Fang, Jie Shi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that MEx MI can improve up to 11.14% from the best known PAME attack and reach 94.07% fidelity with only 16k queries.
Researcher Affiliation Collaboration Yaxin Xiao The Hong Kong Polytechnic University 20034165r@connect.polyu.hk Qingqing Ye The Hong Kong Polytechnic University qqing.ye@polyu.edu.hk Haibo Hu The Hong Kong Polytechnic University haibo.hu@polyu.edu.hk Huadi Zheng The Hong Kong Polytechnic University huadi.zheng@connect.polyu.hk Chengfang Fang Huawei International, Singapore fang.chengfang@huawei.com Jie Shi Huawei International, Singapore shi.jie1@huawei.com
Pseudocode Yes As illustrated in Fig. 1 and pseudocode in Appendix A, the input of MEx MI is an adversary data pool P and the access to a black-box victim model F, and its outputs are the copy model F and the inferred training dataset ˆD.
Open Source Code Yes The codes are available at https://github.com/mexmi/mexmi-project.
Open Datasets Yes Datasets. We perform PAME attacks on two image datasets, namely CIFAR10 [28] and Street View House Number (SVHN) [32], and a text dataset AG’S NEWS which contains corpus of AG’s news articles [11] (see Appendix E.1 for details).
Dataset Splits No The paper mentions training and test datasets (e.g., 'CIFAR10 and AG’S NEWS test sets') but does not specify explicit validation dataset splits or details for reproducibility.
Hardware Specification Yes All experiments are conducted on NVIDIA GeForce RTX 3090, except for those run on Model Arts.
Software Dependencies No The paper mentions using frameworks like Model Arts and specific models (e.g., Wide-Res Net-28-10, DPCNN, VGG16) and optimizers (Adam), but it does not specify version numbers for key software dependencies such as Python, PyTorch, or TensorFlow libraries.
Experiment Setup Yes Recall that the hyper-parameters (such as epoch, initial learning rate, and optimizer) of shadow models Fs can be adjusted to maximize the metric Q. Parameter a in Q is set as 0.05 and f( ) is set as log10( ).3 The preset weights ratio ω in MI Post-Filter is 5 : 1. For CIFAR10 experiments, MEx MI queries 2k samples in each round with a total of 8 rounds. For AG’S NEWS experiments, there are 6 rounds, each with 5k samples.