2Q25 Federated Computing Update

Federated learning (FL) is moving from lab tests to real products. New releases like NVIDIA FLARE in PyTorch mobile, Flower 1.18, and IBM’s privacy toolkit let everyday developers add FL without building custom stacks.

Research this quarter centers on three points. First, communication cuts: quantization and split networks shrink updates to fit narrow links. Second, stronger privacy: differential-privacy methods now ship by default, even for large language models. Third, coping with messy data: adaptive aggregation and meta-learning handle non-IID datasets.

Use cases are expanding. Healthcare studies span from brain-tumor segmentation to pandemic monitoring, while finance teams focus on real-time fraud detection under strict data-residency rules. Energy, IoT security, traffic control, and even agricultural yield prediction showcase FL’s ability to share insights across siloed operators without exposing raw data. The common thread: industries with strict privacy requirements or fragmented data landscapes are embracing cross-organization collaboration to unlock model accuracy that single institutions cannot achieve alone.

Finally, FL is converging with edge and cross-device computing. ExecuTorch plus FLARE points to a near-future where smartphones, wearables, and IoT sensors participate in global model training, enabling personalized experiences and real-time analytics while keeping data local. 

Together, these trends position federated computing as the default architecture for AI where privacy, distribution, and scale intersect. This outlook underscores Rhino’s mission and the value our platform delivers across regulated industries.

Major FL Framework Updates

• NVIDIA and Meta: NVIDIA and Meta’s PyTorch team have announced a major collaboration that brings federated learning capabilities to mobile devices through the integration of NVIDIA FLARE and ExecuTorch. You can now define model architecture and training parameters using familiar PyTorch code and migrate it to cross-device FL paradigm.

• NVIDIA FLARE: A major bottleneck in FL is the exchange of model updates among remote participants and servers, as the size of these messages can be prohibitively large, leading to increased latency and bandwidth consumption. The NVFlare team shares how message quantization offers a solution by reducing the precision of transmitted updates, thereby compressing the message size. Read more here.
• Flower Labs: Flower 1.18.0 is now live, bringing smoother workflows and full Python 3.12 support! Read more here.
• IBM: Ahead of the ICLR 2025 Conference, researchers at IBM propose two new approaches. The first is LASER-VFL, a vertical FL method designed for efficient training and inference of split neural network models that can handle incomplete feature partitions. The second is Differentially Private Federated Prompt Learning (DP-FPL) for multimodal LLMs, a low-rank adaptation scheme with local differential privacy applied to components of the local prompt and global differential privacy to the global prompt.

‍Applications of Federated Computing 

During 2Q25, federated computing entered day-to-day use. Hospitals, banks, utilities, and device makers trained shared models while their raw data stayed put. Results include sharper tumour detection, quicker fraud alerts, smarter power forecasts, and traffic lights that learn local flow. The take-away is simple: when data must remain on-site, federated learning now gives teams a tested path to collective insight.

‍Healthcare

• “Towards fairness-aware and privacy-preserving enhanced collaborative learning for healthcare”: This paper introduces DynamicFL, a resource-adaptive framework for collaborative learning in healthcare that addresses fairness and privacy concerns stemming from varying computational capacities among institutions.
• “FAItH: Federated Analytics and Integrated Differential Privacy with Clustering for Healthcare Monitoring”: This paper introduces FAItH, a dual-stage solution for privacy-preserving healthcare monitoring that integrates federated analytics with differential privacy and clustering. 
• “Leveraging federated learning and edge computing for pandemic-resilient healthcare”: This study presents a pandemic-compliant mechanism for monitoring biosafety protocols like facemask use, social distancing, and contact tracing, integrated with cyber-attack detection. The system, built on IoT, edge computing, and a federated learning framework, achieved high accuracy across these tasks and demonstrates the efficacy of decentralized, AI-driven solutions for public health monitoring in residential settings.
• “Federated learning with integrated attention multiscale model for brain tumor segmentation”: This research proposes a novel FL model, Mixed-FedUNet, for privacy-preserving brain tumor segmentation using the BraTS 2020 dataset. It integrates a Reinforcement Learning-based Federated Averaging (RL-FedAvg) algorithm and a Double Attention-based Multiscale Dense-U-Net, achieving high accuracy (98.24%) and Dice coefficient (93.28%) while maintaining patient data confidentiality.
• “Privacy-preserving federated learning for collaborative medical data mining in multi-institutional settings”: This study integrates transfer learning and federated learning for privacy-preserving medical image classification across multiple institutions, with an adaptive aggregation method that dynamically switches between FedAvg and FedSGD based on data divergence and achieves high accuracy on diverse medical datasets.
• “Federated learning with differential privacy for breast cancer diagnosis enabling secure data sharing and model integrity”: This paper explores the integration of FL and differential privacy for breast cancer detection, enabling collaborative model training across healthcare organizations without exposing raw patient data. The proposed FL-DP model achieves 96.1% accuracy with a privacy budget of ϵ=1.9, validating its feasibility for reliable and privacy-preserving AI in clinical applications.
• “Privacy-preserving Federated Learning and Uncertainty Quantification in Medical Imaging”: Researchers at the Moffitt Cancer Center, University of South Florida, and Rowan University review the latest advancements in federated learning, privacy preservation, and uncertainty quantification in medical imaging. 
• The Danish Technical University is doing a Federated Learning project for personalized audiology to improve hearing aid performance
• UNICEF explores how federated learning protects sensitive children's data in public sectors like health and education without compromising privacy.

‍Financial Services

Recent research here affirms our work with SWIFT to combat payment fraud and our continued focus on this vertical.

• “Optimizing Infrastructure Services in Banking IT with Federated Learning and AI Governance” provides another lens into how FL and AI governance can optimize banking IT infrastructure services, with examples of successful implementation improving credit scoring and detecting fraud without compromising sensitive customer data.
• “Bank data protection and fraud identification based on improved adaptive federated learning and WGAN” proposes an improved adaptive data protection architecture for banks by combining FL and differential privacy, aiming to build a more secure and efficient data privacy protection system. It also designs a fraud recognition model based on Wasserstein Generative Adversarial Networks (WGAN) to effectively identify transaction fraud, demonstrating high accuracy (0.996) and robustness against attacks under low privacy budgets.

‍Security

• “Sentimental analysis based federated learning privacy detection in fake web recommendations using blockchain model” introduces a privacy-focused system for detecting fake web recommendations by integrating sentiment analysis, federated learning, and blockchain technology. The model achieves high accuracy (up to 99%) in identifying fake content, leveraging a generative convolutional Bernoulli Bayes neural network for feature extraction and classification.
• “An intelligent federated learning boosted cyberattack detection system for Denial-Of-Wallet attack using advanced heuristic search with multimodal approaches”: This study proposes the CDMDOW-AMOAFL model, a novel federated learning-based system designed to detect and mitigate Denial-of-Wallet (DoW) cyberattacks. It employs z-score normalization and Harris Hawk Optimization (HHO) for feature selection, combined with an ensemble of GRU, TCN, and CAE models whose hyperparameters are tuned by a Modified Marine Predator Algorithm (MMPA), achieving a superior accuracy of 98.12%.
• “An efficient trustworthy cyberattack defence mechanism system for self guided federated learning framework using attention induced deep convolution neural networks”: This paper presents the CDMFL-AIDCNN technique, an advanced self-guided federated learning framework that uses attention-induced deep convolutional neural networks for robust cyberattack defense in distributed systems. This approach achieved impressive accuracy values of 99.07% on the CIC-IDS-2017 and 98.64% on the UNSW-NB15 datasets, leveraging a hybrid classification model optimized by a Growth Optimizer (GO).
• “LSTM-JSO framework for privacy preserving adaptive intrusion detection in federated IoT networks”: A novel framework proposed in this paper enhances intrusion detection in IoT systems through a decentralized FL approach by dynamically optimizing hyperparameters using the Joint Strategy Optimization (JSO) algorithm. The model achieves superior performance and robustness across various IoT datasets, including 99.5% accuracy on the TON_IoT dataset.
• In this AFCEA International article, federated learning is positioned as bolstering national security. 

‍IoT

• “A framework reforming personalized Internet of Things by federated meta-learning” introduces Cedar, which enables safeguarded knowledge transfer for highly generalizable models that can be rapidly adapted by individuals, significantly improving learning cost, efficiency, speed, and security across various domains while defending against malicious attacks. 
• “UAV-Assisted Multi-Task Federated Learning with Task Knowledge Sharing”: Unmanned Aerial Vehicle (UAV)-assisted networks are seen as an important enabler of post-5G and upcoming 6G networks. Researchers from China propose a novel UAV-assisted multi-task federated learning scheme, where the data collected by the UAVs can be used to train multiple related tasks, and the system performance and robustness can be improved by sharing knowledge among related tasks.

‍Hardware

• “Low Power Ternary State Channel Computing-in-Memory Transistor for Federated Learning”: This paper presents a new ternary state channel computing-in-memory transistor that is designed to enhance federated learning by reducing its communication and energy demands. By leveraging the transistor's ability to distinguish between three conductivity states for weight changes, a custom FL task achieved an 83.3% reduction in total communication bits, showing its potential for low-power ternary computing. 

‍Energy

• “Heuristic based federated learning with adaptive hyperparameter tuning for households energy prediction”: This paper proposes a hierarchical federated learning solution for household energy prediction, tackling non-IID data challenges via clustering and adaptive hyperparameter tuning. It significantly improves prediction accuracy, cuts network traffic below 30 KB, and reduces communication rounds by 30% through optimized aggregation and a genetic algorithm.
• “Federated learning-based non-intrusive load monitoring adaptive to real-world heterogeneities”: This research proposes an FL-NILM that adapts to real-world heterogeneities in electricity consumption data, local models, and AMI facilities. It notably outperforms existing approaches, achieving a 74.7% federated gain in MAE and significantly reducing communication complexity by 95.4%.
• “Energy Minimization for Participatory Federated Learning in IoT Analyzed via Game Theory”: This paper uses game theory to investigate energy minimization in participatory federated learning for IoT, with nodes controlling participation based on local costs. It concludes that a fully distributed optimization is unfeasible without incentives, leading to substantial performance drops (at least 28%) versus centralized solutions.

‍Civil Engineering

• “Federated Machine Learning Enables Risk Management and Privacy Protection in Water Quality”: Real-time water quality risk management in wastewater treatment plants requires extensive data, but data privacy inhibits data sharing. These authors show an adaptive water system federated averaging (AWSFA) framework based on FL that is more robust than direct training and classical federated averaging.
• “Federated deep reinforcement learning-based urban traffic signal optimal control”: Researchers proposed a cross-domain intelligent traffic signal control method using federated Proximal-Policy Optimization (PPO) to address issues like slow learning speed and poor model generalization in deep reinforcement learning for urban traffic signal optimization. This approach significantly improved model generalization and accelerated convergence, with experiments showing reductions in average vehicle waiting time by up to 27.34% and convergence speeds up to 47.69% faster than individual PPO training.

Automotive

• “DRL-Based Resource Allocation for Motion Blur Resistant Federated Self-Supervised Learning in IoV” advances a Deep Reinforcement Learning-based resource allocation scheme, DRL-BFSSL, to minimize energy consumption and latency in a motion blur-resistant federated learning method for the Internet of Vehicles. The solution, which also addresses privacy leakage by using dual temperature instead of a dictionary, validates its effectiveness through simulation results in aggregating models and allocating resources based on motion blur.

‍Agriculture

• “Balancing centralisation and decentralisation in federated learning for Earth Observation-based agricultural predictions”: Crop yield prediction using Earth Observation data is difficult due to diverse data modalities and limited, often private, datasets. The study showed that increasing data aggregation levels (fewer clients with more data) improves learning outcomes by mitigating issues with client numbers, data distribution imbalances, and privacy utility challenges. While increased centralization enhances learning, it involves a trade-off with privacy, suggesting that biome or region-specific models could provide a balanced compromise.

‍Edge Computing

• Edge AI is all about running AI workloads where the data is created and enhancing privacy and latency. CIOs continue to plan their next generation architectures with edge AI capabilities, with companies like Rockwell Automation working closely with NVIDIA. Read more here. 
• Microsoft is advancing edge computing in the defense sector through its Microsoft Adaptive Cloud approach and Azure Local. Read more here. 

‍