Backdoor Federated Learning by Poisoning Backdoor-Critical Layers

Authors: Haomin Zhuang, Mingxian Yu, Hao Wang, Yang Hua, Jian Li, Xu Yuan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our BC layer-aware backdoor attacks can successfully backdoor FL under seven SOTA defenses with only 10% malicious clients and outperform latest backdoor attack methods.
Researcher Affiliation Academia Haomin Zhuang1, Mingxian Yu1 , Hao Wang2, Yang Hua3, Jian Li4, Xu Yuan5 1South China University of Technology 2Louisiana State University 3Queen s University Belfast, UK 4Stony Brook University 5University of Delaware
Pseudocode No The paper describes methods with numbered steps and diagrams (e.g., Section 3, Figure 2), but it does not include formal pseudocode blocks or algorithms labeled as such.
Open Source Code No The paper does not contain any explicit statements about making the source code available, nor does it provide a link to a code repository.
Open Datasets Yes Datasets: Fashion-MNIST (60,000 images for training and 10,000 for testing with ten classes) and CIFAR-10 (50,000 for training and 10,000 for testing with ten classes).FEMNIST is a real-world dataset included in LEAF (Caldas et al., 2018).
Dataset Splits Yes A local dataset D(i) in i-th malicious client is split into training sets D(i) clean,train and D(i) poison,train as well as validation sets D(i) clean,val and D(i) poison,val.
Hardware Specification Yes We conduct all experiments using a NVIDIA RTX A5000 GPU.
Software Dependencies No The paper mentions using 'Py Torch' but does not specify a version number or other software dependencies with their respective versions.
Experiment Setup Yes The proportion of clients selected in each round among n = 100 clients is C = 0.1. Each selected clients train E = 2 epochs in the local dataset with batch size B = 64. The server trains the global model with R = 200 rounds to make it converge. We set τ = 0.95 when identifying the BC layers via Layer Substitution Analysis... We set λ = 1 when training on CIFAR-10 and λ = 0.5 when training on Fashion-MNIST... Table A-7 in Appendix shows the detailed hyperparameter settings.