Hi-SAFE: Hierarchical Secure Aggregation for Lightweight Federated Learning

A new decentralized peer-to-peer framework for federated learning (FL) has been proposed, challenging the traditional centralized star topology that limits personalization and robustness. This innovative approach allows clients to aggregate personalized updates from trusted peers, enhancing model training while maintaining data privacy.

A novel framework called FedNCA-ML has been proposed to enhance Federated Learning (FL) in multi-label scenarios, specifically addressing the challenges posed by label-distribution skew. This framework aligns feature distributions across clients and learns well-clustered representations inspired by Neural Collapse theory, which is crucial for applications like medical imaging where data privacy and heterogeneous distributions are significant concerns.

A new analytical prediction algorithm has been proposed for edge computing environments in agriculture, aimed at reducing the excessive transmission of sensor data generated by IoT devices. This approach utilizes a predictive filter to forecast sensor readings, triggering data transmission only when significant deviations occur, thereby addressing issues of network congestion and energy consumption in resource-constrained settings.

A novel generative AI-powered plugin has been proposed to enhance federated learning in heterogeneous IoT networks, addressing the challenges posed by Non-IID data distributions that hinder model convergence. This approach utilizes generative AI for data augmentation and a balanced sampling strategy to synthesize additional data for underrepresented classes, thereby improving the robustness and performance of the global model.

A recent perspective paper highlights the vulnerabilities of Federated Learning (FL) in military applications, particularly concerning Large Language Models (LLMs). It identifies prompt injection attacks as a significant threat that could compromise operational security and trust among allies. The paper outlines four key vulnerabilities: secret data leakage, free-rider exploitation, system disruption, and misinformation spread.

The introduction of pFedBBN, a personalized federated test-time adaptation framework, addresses the critical challenge of class imbalance in federated learning. This framework utilizes balanced batch normalization to enhance local client adaptation, particularly in scenarios with unseen data distributions and domain shifts.

A novel Q-learning framework has been proposed for time-critical data aggregation scheduling in Internet of Things (IoT) networks, aiming to reduce latency in applications such as smart cities and industrial automation. This approach integrates aggregation tree construction and scheduling into a unified model, enhancing efficiency and scalability.

A new study has introduced a problem-oriented taxonomy of evaluation metrics for time series anomaly detection, addressing the challenges posed by diverse application objectives and varying metric assumptions in IoT and cyber-physical systems. The framework categorizes over twenty metrics into six dimensions, focusing on their specific evaluation challenges rather than their mathematical forms. Comprehensive experiments were conducted to assess metric behavior across different detection scenarios.