92 papers - avg viability 5.6
Federated Learning (FL) is advancing the field of decentralized machine learning by enabling collaborative model training without compromising data privacy. Current research focuses on addressing challenges such as domain shifts, client heterogeneity, and communication efficiency. Innovations like adaptive learning frameworks and prototype-based methods are enhancing model performance across diverse client data while minimizing resource consumption. These developments are crucial for builders aiming to implement scalable and privacy-preserving AI solutions in real-world applications, particularly in sensitive domains like healthcare and finance. By leveraging federated learning, organizations can harness the power of distributed data while maintaining compliance with privacy regulations, ultimately leading to more robust and generalized models.
FedBPrompt enhances federated person re-identification by using learnable visual prompts to improve feature discrimination across decentralized data.
FairFAL is an adaptive federated active learning framework that enhances performance in class-imbalanced and non-IID settings.
HeteroFedSyn is a framework for differentially private tabular data synthesis in heterogeneous federated settings, enabling secure data sharing for various tasks.
QuantFL is a sustainable federated learning framework that reduces energy costs for IoT devices through efficient model quantization.
REED is a novel noncoherent aggregation primitive for over-the-air federated learning that avoids instantaneous CSI, enabling stable convergence even with data heterogeneity.
FedMPT is the first method for federated multi-label recognition that uses LLM-driven pipelines and optimal transport to mitigate erroneous label activations and improve model robustness.
PromptGate enhances federated active learning by dynamically filtering out-of-distribution noise using a vision-language model, improving annotation efficiency in medical AI.
A federated learning framework for aligning heterogeneous vision-language models using preference-based collaboration, enabling privacy-preserving model updates without direct data or parameter sharing.
EASE is a framework for federated multimodal unlearning that addresses entanglement across modalities and client subspaces.
A data-free method for estimating client contribution in federated learning using the spectral entropy of final-layer updates, enabling privacy-preserving and fair reward mechanisms.