High-Fidelity Gradient Inversion in Distributed Learning

Authors: Zipeng Ye, Wenjian Luo, Qi Zhou, Yubo Tang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate the superiority of our approach, reveal the potential vulnerabilities of the distributed learning paradigm, and emphasize the necessity of developing more secure mechanisms.
Researcher Affiliation Academia Zipeng Ye1, 2, Wenjian Luo1, 2, 3*, Qi Zhou1, 2, Yubo Tang1, 2 1School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen 2Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies 3Peng Cheng Laboratory
Pseudocode Yes Pseudocode of this part is provided in Appendix A.
Open Source Code Yes Source code is available at https://github.com/Mi Lab-HITSZ/2023Ye HFGrad Inv.
Open Datasets Yes We conduct experiments for large-scale image classification task using Image Net ILSVRC 2012 dataset (Deng et al. 2009), as well as randomly collected images from Web.
Dataset Splits Yes We randomly select 64 images from the validation set of Image Net, and their label distributions are shown in Fig. 4 (a).
Hardware Specification Yes optimized with 15K iterations on NVIDIA TITAN RTX GPUs.
Software Dependencies No The paper mentions using PyTorch library and Adam optimizer, but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes The dropout rate we used in (10) is set as 0.3. We set β = 2 in (11), ˆλtv = 0.01, ˆλBN = 10 4, T1 = 3, 000, T2 = 5, 000 in scheduling strategy. We use Adam (Kingma and Ba 2014) for optimization with a step learning rate decay, and each batch is optimized with 15K iterations on NVIDIA TITAN RTX GPUs.