UMEDA: Unified Multi-modal Efficient Data Fusion for Privacy-Preserving Graph Federated Learning
Researchers have introduced UMEDA, a novel graph federated learning framework designed to enhance device-free localization using heterogeneous sensors like Wi-Fi and LiDAR. Traditional federated learning struggles with varying sensor modalities, data distribution drift, and privacy noise that degrades structural signals. UMEDA addresses these challenges by modeling clients as nodes in a global graph sharing a continuous integral operator, where aggregation is treated as spectral signal processing. The system employs linear-attention layers with low-rank filtered kernel spectra to align diverse sensors into a common subspace, suppressing modality-specific residuals. Server-side aggregation utilizes a diffusion model on spectral coefficients, allowing flexibility in graph sizes and missing modalities without requiring node-wise correspondence. To balance privacy and utility, an anisotropic differential-privacy mechanism projects noise into the null space of the signal subspace, preserving key eigendirections while ensuring formal differential privacy. Benchmarks on MM-Fi and RELI11D datasets demonstrate that UMEDA significantly outperforms state-of-the-art baselines in accuracy, convergence speed, and communication efficiency, particularly under conditions of high modality heterogeneity and strict privacy constraints.
Wire timeline
UMEDA: Unified Multi-modal Efficient Data Fusion for Privacy-Preserving Graph Federated Learning
Researchers have introduced UMEDA, a novel graph federated learning framework designed to enhance device-free localization using heterogeneous sensors like Wi-Fi and LiDAR. Traditional federated learning struggles with varying sensor modalities, data distribution drift, and privacy noise that degrades structural signals. UMEDA addresses these challenges by modeling clients as nodes in a global graph sharing a continuous integral operator, where aggregation is treated as spectral signal processing. The system employs linear-attention layers with low-rank filtered kernel spectra to align diverse sensors into a common subspace, suppressing modality-specific residuals. Server-side aggregation utilizes a diffusion model on spectral coefficients, allowing flexibility in graph sizes and missing modalities without requiring node-wise correspondence. To balance privacy and utility, an anisotropic differential-privacy mechanism projects noise into the null space of the signal subspace, preserving key eigendirections while ensuring formal differential privacy. Benchmarks on MM-Fi and RELI11D datasets demonstrate that UMEDA significantly outperforms state-of-the-art baselines in accuracy, convergence speed, and communication efficiency, particularly under conditions of high modality heterogeneity and strict privacy constraints.
cs.AI updates on arXiv.org