Energy-based Backdoor Defense Against Federated Graph Learning

Authors: Guancheng Wan, Zitong Shi, Wenke Huang, Guibin Zhang, Dacheng Tao, Mang Ye

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive results on various settings of federated graph scenarios under backdoor attacks validate the effectiveness of this approach. We conducted experiments on five mainstream datasets under both IID and Non-IID scenarios, as well as with varying proportions of malicious clients. The results demonstrate that our approach outperforms the current state-of-the-art methods in traditional FL. We plotted the curves of V during the training process on the Pubmed dataset with a malicious proportion of Υ = 0.3 and 0.5. We conducted an ablation study on the key components of our method using the Cora and Pub Med datasets.
Researcher Affiliation Academia 1 National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University 2 Taikang Center for Life and Medical Sciences, Wuhan University 3 Tongji University 4 Generative AI Lab, College of Computing and Data Science, Nanyang Technological University
Pseudocode Yes Algorithm 1 Fed TGE Input: Communication rounds T, participant scale K, kth client private model wk, and local data Gk Output: The final global model MT for t = 1, 2, , T do Client Side: for k = 1 to K in parallel do fk( ) Local Updating(wt, Gk) // Original training strategy f t k( ) Energy Calibrating(fk( ), Gk) // Injecting distribution knowledge Et k = {E(f t k(vi))}Nk i=1 // Calculating energy elements Server Side: // Cluster and find the cluster with the smaller mean {Et a, Et b} and {Et k}N k =a,b where mean(Et a, Et b) mean({Et k}N k =a,b) St = {St mn}N mn = cos(Et m, Et n)m,n =a,b // Calculating energy elements similarity Ge = (Ve, Ee) (τ, St) // Constructing energy graph Et k (βt k, {Et k}N k =a,b) // Energy graph similarity propagation It k (βt k, {Et k }N k =a,b) // Energy disparity aggregation wt+1 = PN l=1 Ikwt k // Model parameter update
Open Source Code Yes The code is available at https://github.com/Zitong Shi/fed TGE.
Open Datasets Yes We evaluate the efficacy and robustness in three scenarios: Citation Network (Yang et al., 2016a), Co-authorship (Shchur et al., 2018), and Amz-purchase (Mc Auley et al., 2015). Detailed information about the datasets is provided in appendix A. Table 4: Statistics of datasets used in experiments. Dataset #Nodes #Edges #Classes #Features Cora 2,708 5,278 7 1,433 Pubmed 19,717 44,324 3 500 Coauthor-CS 18,333 327,576 15 6,805 Amz-Physics 34,493 495,924 5 8415 Amz-Photo 7,650 287,326 8 745
Dataset Splits Yes Dataset Split: In this paper, we conduct experiments on the node classification task. Following (Xu et al., 2021; Zheng et al., 2023), we partition the original training dataset (labeled) into training, validation, testing sets, comprising 60%, 20%, and 20% of the total nodes, respectively.
Hardware Specification No The supercomputing system at the Supercomputing Center of Wuhan University supported the numerical calculations in this paper. No specific hardware models (e.g., GPU/CPU models, memory details) are mentioned.
Software Dependencies No The Adam optimizer (Kingma, 2014) with a learning rate of 0.01 is used to train the GNN models. No specific version numbers for software libraries (e.g., PyTorch, TensorFlow) or programming languages (e.g., Python) are provided.
Experiment Setup Yes Training Setting: We repeat each experiment five times for each federated approaches to ensure the robustness and reliability of the results. The Adam optimizer (Kingma, 2014) with a learning rate of 0.01 is used to train the GNN models. Network Structure: Following the common approach in FGL(Dai et al., 2023), we utilize GCN as the 2 layers feature extractor and classifier, with the hidden layer size of 32 for all datasets. Backdoor Attack: We set the malicious client ratio Υ as {0.1, 0.3, 0.5}. Hyper-Parameters: We observed that the choice of hyperparameters does not lead to significant fluctuations in the evaluation metrics. In most of our experiments, we set the default values as 10 for energy-epochs and 0.8 for τ.