Dual Defense: Enhancing Privacy and Mitigating Poisoning Attacks in Federated Learning

Authors: Runhua Xu, Shiqi Gao, Chao Li, James Joshi, Jianxin Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted extensive experiments on various model poisoning attacks and FL scenarios, including both cross-device and cross-silo FL. Experiments on publicly available datasets demonstrate that DDFed successfully protects model privacy and effectively defends against model poisoning threats.
Researcher Affiliation Academia Runhua Xu Beihang University runhua@buaa.edu.cn Shiqi Gao Beihang University gaoshiqi@buaa.edu.cn Chao Li Beijing Jiaotong University li.chao@bjtu.edu.cn James Joshi University of Pittsburgh jjoshi@pitt.edu Jianxin Li Beihang University and Zhongguancun Laboratory lijx@buaa.edu.cn
Pseudocode Yes Due to space limitations, the formal algorithm pseudocode is provided solely in Appendix A.1.
Open Source Code Yes The experimental DDFed is available on the Git Hub repository.
Open Datasets Yes We assessed our proposed DDFed framework using publicly available benchmark datasets: MNIST[19], a collection of handwritten digits, and Fashion-MNIST (FMNIST)[33], which includes images of various clothing items, offering a more challenging and diverse dataset for federated learning tasks.
Dataset Splits No The paper mentions using MNIST and FMNIST datasets but does not explicitly state the training/validation/test dataset splits needed for reproduction. It mentions creating
Hardware Specification Yes Note that the time-related experiments were conducted on a Mac OS platform with an Apple M2 Max chip and 96GB of memory.
Software Dependencies No This secure aggregation is implemented through Ten SEAL library [4].
Experiment Setup Yes The default FL training involves 10 clients randomly chosen from 100 for each communication round. Furthermore, we employ a batch size of 64 with each client conducting local training over three epochs per round using an SGD optimizer with a momentum of 0.9 and a learning rate of 0.01. Our DDFed implementation s default epsilon (ε) value is set to 0.01 unless specified differently.