Proactive Privacy-preserving Learning for Retrieval

Authors: Peng-Fei Zhang, Zi Huang, Xin-Shun Xu3369-3376

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three widely-used realistic datasets prove the effectiveness of the proposed method.
Researcher Affiliation Academia Peng-Fei Zhang,1 Zi Huang, 1 Xin-Shun Xu 2 1 School of Information Technology & Electrical Engineering, University of Queensland, Brisbane, Australia 2 School of Software, Shandong University, Jinan, China
Pseudocode Yes Algorithm 1 Proactive Privacy-preserving Learning Input: Training data: X, mini-batch size: m, learning rate: σg, σf, hash code length: c, iteration times: t, epoch times: e = [t/m] Output: Adversarial generator: G( ; ϑg) Procedure: Randomly initialize ϑg, ϑf for i = 1 : t do for j = 1 : e do 1. xi X // Construct a mini-batch 2. ϑg ϑg σg θg(Jo Jp) // Update ϑg by SGD 3. ϑf ϑf σf θf Jp // Update ϑf by SGD end for end for Return: ϑg
Open Source Code No The paper does not contain an explicit statement about releasing source code for the described methodology or a link to a code repository.
Open Datasets Yes To validate the effectiveness of the proposed PPL, we conduct extensive experiments on three widelyused benchmark datasets, i.e., CIFAR-10 (Krizhevsky et al. 2009), Image Net (Deng et al. 2009), and Fashion MNIST (Xiao, Rasul, and Vollgraf 2017).
Dataset Splits No The paper specifies 10,000 data points randomly selected for training and some images for testing (e.g., 1000 for CIFAR-10, 1300 for ImageNet, 10,000 for Fashion MNIST test set), but does not explicitly mention a separate 'validation' split.
Hardware Specification Yes The experiments are implemented with Pytorch on a workstation (with Intel XEON E5-2650 v3 @ 2.60GHz CPU, NVIDIA 1080Ti GPU).
Software Dependencies No The paper mentions 'Pytorch' but does not specify its version or other software dependencies with version numbers.
Experiment Setup Yes Implementation Details For experiments on CIFAR-10 and Image Net, Res Net50 (He et al. 2016) network is adopted as the backbone of the surrogate retrieval model... The parameters are set as follows. we set α = γ = λ = 1, β = 10. The perturbation constraints are selected as the l norm with ϵ as 0.1,0.032,0.3 for CIFAR-10, Image Net and Fashion MNIST, respectively. The generated images would be projected into [0, 1] space. The learning rates are 10 4 and 10 3 for the generator and the retrieval model. The iterations are respectively fixed as 500,500 and 1000. We fix the batch sizes as 32,32,512.