Research map / Federated learning and privacy / Privacy attacks and secure learning
EBS-CFL: Efficient and Byzantine-robust Secure Clustered Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EBS-CFL, a single-server secure aggregation scheme for clustered federated learning that hides client gradients and cluster identities via orthogonal-matrix encoding, filters Byzantine gradients with a secure ReLU on cosine similarity, and adds compression for efficiency.
Problem
Clustered federated learning under non-i.i.d. data needs to protect clients' gradients and cluster identities from the server while remaining robust to Byzantine attacks and efficient, with only a single cloud server.
Approach
Combines IFCA-style clustered aggregation with a cosine-similarity-based robust weighting (RFCA), realized under encryption by VOMCA (orthogonal matrix confusion with verifiable encodings), SRFC (secure ReLU via squaring/square-root tricks), and compression (gradient segmentation and layered aggregation).
Key results
- Under 40% Sine and LF attacks with Dirichlet alpha=0.5, EBS-CFL keeps final accuracy about 65% (LF) and 65% (Sine) while FLTrust drops to ~57% and IFCA/FedAvg are heavily hurt in some settings (Table 2).
- On MNIST with five attack types, EBS-CFL reports 97.82% (LF) and 98.43% (Adaptive), above FLTrust's 96.82% and 97.76% (appendix table).
- Client-side computation is O(m^2 l) and per-client communication is O(ml + m^2); with m=1 client cost is claimed at least O(log n) times better than comparison schemes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EBS-CFL instead of FLTrust | Byzantine-robust federated learning · MNIST, Byzantine attacks | accuracy | Mixed |
| Use EBS-CFL instead of FLTrust | Byzantine-robust federated learning · CIFAR10, Byzantine attacks | accuracy | Helped |
| Use EBS-CFL instead of FLTrust | Byzantine-robust federated learning · Non-i.i.d., LF and Sine attacks, α=0.1, 0.5, and 0.9 | final accuracy | Helped |
| Use EBS-CFL instead of FedAvg | Clustered federated learning · MNIST and CIFAR10 with uneven data distributions | accuracy | Helped |
| Use IFCA instead of FedAvg | Non-i.i.d. federated learning under LF attack · CIFAR100, α=0.1 | accuracy | Helped |
| Use IFCA instead of FedAvg | Non-i.i.d. federated learning under LF attack · CIFAR100, α≤0.5 | accuracy | Hurt |
| Adjust training of the server update g0 (ablation) | EBS-CFL · Model convergence experiments | accuracy | Helped |
| Remove the ||g0|| scaling factor (its effect can instead be absorbed into the learning rate) (ablation) | EBS-CFL · Model convergence experiments | accuracy | No clear effect |
Surprising findings
- Accuracy under the Sine attack at alpha=0.5 for IFCA is lower than FedAvg-level behaviour, and the authors note LF attacks hurt IFCA more than FedAvg.
Limitations
- Accuracy drops notably at extreme heterogeneity (alpha=0.1), which the authors attribute to heterogeneity rather than attackers.
- Security relies on a single cloud server and a KDC; the proofs are heuristic (e.g., rotation argument and probabilistic guessing) and the theorems are loosely stated.
- Server-side computation grows as O(m^2 n^2 + nm^3 l + nm^4) and depends on the server update g_0, which the paper notes affects convergence.
Experimental setup
Datasets: MNIST CIFAR10 CIFAR100
Benchmarks: MNIST CIFAR10 CIFAR100
Metrics: final accuracy maximum accuracy attack success rate attack impact rate communication overhead computation time
Baselines: FedAvg FLTrust IFCA Median Krum CCESA FastSecAgg SwiftAgg+
Builds on: IFCA FLTrust SecAgg Sine Attack
Closest papers by meaning
- Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.90)
- Efficient Federated Learning against Byzantine Attacks and Data Heterogeneity via Aggregating Normalized Gradients (NeurIPS 2025 · similarity 0.87)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.86)
- Towards Trustworthy Federated Learning with Untrusted Participants (ICML 2025 · similarity 0.86)
- Self-Driven Entropy Aggregation for Byzantine-Robust Heterogeneous Federated Learning (ICML 2024 · similarity 0.86)
- Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.