A new arXiv paper proposes jointly optimizing federated learning and wireless transmission to counter client drift caused by heterogeneous data and multiple local updates. The approach also aims to handle unreliable wireless links, where transmission errors can invalidate model updates.
A new arXiv paper presents a federated learning framework designed to make collaborative training across distributed data safer. It combines dynamic differential privacy, a lightweight homomorphic encryption scheme, and asynchronous aggregation to reduce privacy risks while limiting computational overhead. The work is posted as a preprint and has not yet been peer reviewed.
A new arXiv preprint examines integrity risks in federated learning, where an aggregator coordinates training across parties without pooling raw data. The authors argue that once participants or infrastructure cannot be fully trusted, additional guarantees are needed, and they outline two requirements their approach aims to satisfy. The work targets end-to-end verification alongside robustness for the federated setting.
A new arXiv paper proposes pruning visual tokens based on the question being asked, rather than sending all visual features to the server in split-learning setups. The approach targets vision-language question answering in federated, split, and U-shaped split learning, where raw data stays local but transmitted representations can still leak information. The authors frame selective transmission as a way to reduce both privacy exposure and bandwidth use.
A new arXiv paper benchmarks four widely used federated learning frameworks — FedML, Flower, Substra and OpenFL — under a shared experimental setup. The authors assess how each handles scaling and performance, aiming to give practitioners a clearer basis for choosing a framework. The work is a comparative, cross-validated analysis rather than a new model or tool release.
A new arXiv preprint introduces a federated fine-tuning method designed for mixture-of-experts large language models. The approach aims to adapt these sparse-activation models to specific tasks while keeping training distributed. The abstract frames the work as addressing efficiency and capacity trade-offs in MoE architectures.
A revised arXiv paper proposes a federated learning defense framework for synthetic aperture radar target recognition that adapts dynamically to noise. The approach aims to reduce reliance on centralized training, which raises privacy and bandwidth concerns. It is positioned as a cross-domain replacement for an earlier submission.
A new arXiv paper introduces OCT-FedSIR, a federated learning approach for ophthalmic imaging that does not require patient data to be centralized. The work targets unreliable annotation labels across participating institutions, where differences in disease prevalence and class composition can degrade model trustworthiness. It is cross-listed under cs.AI and cs.LG.
A new arXiv preprint examines knowledge graph question answering in settings where no single organization holds the entire graph. The authors describe a scenario in which parties share entity identifiers but own disjoint sets of relations, so federated methods are needed to reason across them. The paper presents design lessons from a working system and an interactive prototype.
The paper compares two federated learning setups: parallel training, where clients work simultaneously, and sequential training, where clients update the model one after another. It derives convergence guarantees for the sequential approach when client data distributions differ, a setting where parallel methods often struggle. The analysis aims to clarify when sequential federated learning offers theoretical advantages over its parallel counterpart.
A new arXiv preprint examines individual rationality in federated learning, where clients weigh privacy, communication, and computation costs against possible model quality gains. The authors introduce a pathwise notion of individual rationality and co-design incentive mechanisms together with system architecture. The work aims to keep participation worthwhile for clients while preserving training outcomes.
A new arXiv paper introduces FLoKD, a method for fine-tuning large language models across wireless networks without centralizing user data. It combines federated learning with low-rank adaptation and adaptive knowledge distillation to reduce communication and computation costs. The approach targets privacy-preserving deployment of LLMs in distributed, bandwidth-limited settings.
A new arXiv paper examines how to audit user-level privacy guarantees in Private Evolution, a method for generating synthetic data in federated settings. The approach collects clipped user votes over a shared candidate bank and turns them into a differentially private histogram with calibrated noise. The work focuses on verifying that individual users' raw data remains protected under this mechanism.
Researchers introduced FedLTLib, a benchmark suite for federated learning in settings where data across clients follows a long-tailed distribution. The work addresses real-world mobile and edge deployments, where privacy constraints keep data decentralized and class frequencies are highly uneven. The benchmark aims to standardize evaluation of methods designed for this combination of challenges.
A new arXiv paper introduces Sylvas, a scheduling method for federated continual learning that selects which devices contribute updates based on their estimated learning value. The approach targets distributed, non-stationary data streams in Internet of Things settings such as intelligent transportation and industrial monitoring. The work appears under both cs.AI (cross) and cs.LG (new) listings as arXiv:2609.15763v1.
A new arXiv paper proposes adaptively switching between training phases to reduce the communication overhead of federated fine-tuning with low-rank adaptation. The authors argue that existing accounting methods for federated LoRA protocols overlook asymmetric transit costs between clients and the server. Their approach aims to make the dominant per-round communication expense more efficient while keeping trainable parameters small on each client.
A new arXiv paper examines how record-level differential privacy clashes with personalized federated learning, where each client's individual variation is low-dimensional but repeated training rounds broadcast high-dimensional model updates. The authors propose a compact adaptation approach paired with variable-length Gaussian communication to reduce that mismatch. The work targets more efficient and privacy-preserving personalization across federated clients.
A new arXiv preprint proposes applying lifecycle cost analysis to federated learning marketplaces that use smart contracts to coordinate training among parties who do not trust each other. The author argues that existing architectures are usually assessed only in isolated economic terms, leaving their full cost profile unclear. The paper aims to provide a more complete framework for evaluating the economics of such blockchain-based FL systems.
A paper summarizes the FedSurg EndoVis 2024 Challenge, which tested federated learning methods on surgical video for appendicitis classification without centralizing patient data. The work addresses the difficulty of building generalizable surgical AI when hospitals cannot share video directly, and reports benchmark outcomes from participating teams. It positions federated training as a viable approach for privacy-sensitive, spatiotemporal surgical tasks.
Researchers propose Block-Wise Marginal Utility Contribution (MUC), a scheme for federated learning that decides which parts of a model update are worth transmitting. By estimating the marginal utility of each block, the framework aims to reduce the communication overhead that typically limits distributed training. The work is published as an arXiv preprint.
A new arXiv preprint addresses a problem in federated learning where clients that run multiple local optimization steps produce parameter updates of very different scales. The authors observe that the proximal term used by FedProx exerts a restraining force that scales only linearly with update size, giving it weak leverage over large displacements, and they propose a higher-order regularization scheme as a stronger alternative.
A new arXiv paper introduces an information-theoretically secure aggregation scheme aimed at federated learning on resource-constrained devices such as smartphones and IoT nodes. The method builds on sign-based gradient compression to keep communication costs low while staying robust to participant dropouts and adversarial behavior.
A recently posted arXiv preprint presents OmniMed-FL, a federated learning framework that combines medical imaging with patient record data for diagnostic tasks. The design keeps model training distributed across institutions rather than centralized, aiming to accommodate privacy rules such as HIPAA and GDPR. The paper appears in both the machine learning and artificial intelligence listings.
A new arXiv paper addresses fine-tuning large language models in settings where participants cannot host the full model locally, often because the model owner keeps it proprietary. The authors introduce a split learning approach designed to preserve privacy during federated fine-tuning on domain-specific data.
A revised arXiv preprint introduces a personalized federated learning method that weights each client's parameter updates according to its influence, rather than relying on fixed aggregation weights. The approach is designed to let devices with different data distributions and preferences train collaboratively while keeping their data private. The updated version appears across arXiv's cs.AI and cs.LG listings.
A new arXiv paper introduces a gradient inversion attack that reconstructs clients' training data from the model updates shared in federated learning. Inspired by LT fountain codes, the method cascades analytic reconstructions through a peeling-style decoding, extending such attacks to larger batches where earlier closed-form inversion approaches break down. The findings underscore that sharing gradients rather than raw data does not fully protect client privacy.
A new preprint presents a federated learning framework that trains biological aging clock models across multiple research centers without centralizing sensitive data, relying on a trust network to coordinate participants. The work also examines which protein interactions, ranging from pairwise to higher-order, contribute most to accurate aging predictions.
A new arXiv paper introduces HybridFLow, a system that uses software-defined networking to decide how to partition clients in cross-silo federated learning. It targets wide-area deployments where network delays dominate the time needed to finish each training round. The approach aims to help distributed institutions train shared models without moving raw data while reducing round completion times.
A new research paper proposes a federated learning framework that separates modality-specific processing so heterogeneous robots in 6G networks can train collaboratively without sharing raw sensor data. The approach targets privacy preservation for embodied AI applications built on low-latency edge connectivity and distributed sensing.
Researchers present NEXUS-MI, an approach that personalizes EEG-based motor-imagery brain-computer interfaces through federated learning, letting user models improve without centralizing sensitive neural recordings. A gateway coordinates training in a communication-aware way, addressing variability across subjects and sessions. The goal is reliable adaptation when each person has only a small amount of calibration data.