Towards Trustworthy Federated Learning with Untrusted Participants

Authors: Youssef Allouah, Rachid Guerraoui, John Stephan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on standard benchmarks validate CAFCOR s practicality, showing that privacy and robustness can coexist in distributed systems without sacrificing utility or trusting the server.
Researcher Affiliation Academia 1EPFL, Switzerland. Alphabetical order. Correspondence to: Youssef Allouah <EMAIL>, John Stephan <EMAIL>.
Pseudocode Yes Algorithm 1 CAFCOR Input: Initial model θ0; DP noise levels σind, σcor; batch size b; clipping threshold C; learning rates {γt}; momentum coefficients {βt}; number of iterations T. Algorithm 2 CAF: Covariance bound-Agnostic Filter Input: vectors x1, . . . , xn Rd; bound on number of corrupt inputs 0 f < n
Open Source Code No To facilitate reproducibility, we intend to publicly release our code.
Open Datasets Yes We consider two widely used image classification datasets: MNIST (Le Cun & Cortes, 2010) and Fashion-MNIST (Xiao et al., 2017).
Dataset Splits Yes On MNIST, we use batch size b = 50, learning rate γ = 0.075, momentum parameter β = 0.85, and clipping parameter C = 2.25. For Fashion-MNIST, we use b = 100, γ = 0.3, β = 0.9, and C = 1. For both datasets, we train for T = 30 iterations and apply ℓ2-regularization at 10 4. We adopt user-level DP across all threat models.
Hardware Specification No The paper does not explicitly describe the specific hardware used (e.g., GPU/CPU models, memory details). It only mentions general concepts like 'distributed environment'.
Software Dependencies No To estimate the privacy budgets achieved at the end of training, we use Opacus (Yousefpour et al., 2021). The citation mentions 'User-friendly differential privacy library in pytorch, 2021', implying PyTorch but no specific version number for PyTorch or Opacus is provided.
Experiment Setup Yes On MNIST, we use batch size b = 50, learning rate γ = 0.075, momentum parameter β = 0.85, and clipping parameter C = 2.25. For Fashion-MNIST, we use b = 100, γ = 0.3, β = 0.9, and C = 1. For both datasets, we train for T = 30 iterations and apply ℓ2-regularization at 10 4. We adopt user-level DP across all threat models. On MNIST, the privacy budgets reach ε = 26.4 and 27.8 for f = 10 and f = 5, respectively. On Fashion-MNIST, the privacy budget is ε = 39.6 for both values of f. Throughout, we set δ = 10 4.