Practical Privacy-Preserving MLaaS: When Compressive Sensing Meets Generative Networks

Authors: Jia Wang, Wuqiang Su, Zushu Huang, Jie Chen, Chengwen Luo, Jianqiang Li

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results confirmed its performance superiority in accuracy and resource consumption against state-of-the-art privacy-preserving MLaa S frameworks. In our experiments, we use three image datasets to evaluate the performance of our model: the MNIST dataset of handwritten digits (Lecun et al. 1998), the Street View House Numbers (SVHN) dataset (Netzer et al. 2011), and the CIFAR-10 dataset (Krizhevsky, Hinton et al. 2009).
Researcher Affiliation Academia Jia Wang1, Wuqiang Su1, Zushu Huang1, Jie Chen 1, Chengwen Luo2, Jianqiang Li2 * 1 College of Computer Science and Software Engineering, Shenzhen University 2 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University 3688 Nanhai Avenue, Shenzhen, Guangdong Province, China {jia.wang, suwuqiang2019}@szu.edu.cn, 2060271076@email.szu.edu.cn, {chenjie, chengwen, lijq}@szu.edu.cn
Pseudocode Yes Algorithm 1: Algorithm of adding noise to original signals of the training dataset
Open Source Code No More details of the hyper-parameters can be found in the code repository. (No specific link or explicit statement of code availability for their work is provided.)
Open Datasets Yes In our experiments, we use three image datasets to evaluate the performance of our model: the MNIST dataset of handwritten digits (Lecun et al. 1998), the Street View House Numbers (SVHN) dataset (Netzer et al. 2011), and the CIFAR-10 dataset (Krizhevsky, Hinton et al. 2009).
Dataset Splits Yes While the test set of all datasets is kept the same, the MNIST, SVHN and CIFAR-10 datasets took out 10,000, 10,000 and 5,000 images from their respective training sets as their respective validation sets.
Hardware Specification No The paper does not provide specific hardware details for running its experiments.
Software Dependencies No Our implementation is based on Tensor Flow (no version specified).
Experiment Setup Yes In all experiments, the measurement matrix is chosen to be a Gaussian random matrix and the Adam optimizer (Kingma and Ba 2015) is used to train the model. In all experiments of DCMG-O and DCMGN, we set λ = 100, 000. In all experiments of DCMGN, for making training dataset we set λnoise = 2, p = 0.5, c = n m for MNIST and λnoise = 1, p = 0.5, c = n m for SVHN and CIFAR-10.