Overview of Research Progress in Federated Learning Technology (2023-2025)
Introduction: Current Status of Federated Learning Technology
Federated Learning (FL), as an important branch of distributed machine learning, has made significant progress in recent years driven by the demand for privacy protection. This paper systematically reviews representative research achievements in this field from 2023 to 2025, covering various application scenarios such as medical image analysis, financial risk control, and smart home systems. Current research mainly focuses on three core challenges: handling data heterogeneity, enhancing privacy security, and optimizing communication efficiency. Notably, with the global implementation of regulations like the General Data Protection Regulation (GDPR), federated learning technology is extending from pure algorithm innovation to compliance framework design—a trend that has become particularly evident in literature over the past two years.
Solutions for Label Heterogeneity in Supervised Federated Learning
Innovation of the Federated Ark+ Architecture
The Federated Ark+ framework proposed at the 2025 Medical Image and Deep Learning Conference (MIDL) systematically addresses long-standing label heterogeneity issues within the medical field. A typical manifestation of this problem is significant differences among different medical institutions regarding annotation standards for identical medical images—such as inconsistencies in tumor boundary delineation or pathological grading criteria. Traditional federated learning methods like FedAvg often experience a decline in model performance under these circumstances because simple parameter averaging blurs boundaries between different labeling systems.
The research team innovatively designed a multi-task head separation architecture where each client retains an independent classifier module while only sharing feature encoder parameters. This design effectively isolates negative impacts arising from label distribution differences while maintaining model generalization capabilities. In experimental phases using a breast mammography dataset collected from six medical centers across North America, Europe, and Asia under similar privacy budget conditions, their classification F1-score improved by 12.7% compared to baseline methods. Particularly noteworthy is that researchers proposed a cyclic training strategy which alternately updates global encoders and local classifiers so that models can maintain personalized adaptation abilities while still absorbing common feature representations across institutions.
Defense Against Byzantine Attacks Under Decentralized Architectures
Robustness Breakthroughs with Balance Aggregation Rules
This study accepted into ACM SIGSAC 2024 re-examines security assumptions within decentralized federated learning environments. Unlike traditional architectures relying on central servers—which avoid single points of failure—fully peer-to-peer network topologies provide more attack surfaces for Byzantine nodes (i.e., malicious participants). The research team theoretically demonstrated that when malicious node proportions exceed 15%, existing detection methods based on cosine similarity will completely fail.
The proposed Balance aggregation rule includes three key innovations: first, it introduces a dynamic weight allocation mechanism based on local test set accuracy allowing normal node updates to gain exponentially increasing aggregation weights; second, it designs lightweight model fingerprint comparison algorithms capable of identifying abnormal parameter fluctuations with only an additional 3% computational overhead; finally, it develops adaptive topology pruning protocols capable of automatically isolating nodes continuously providing low-quality updates. In simulated financial fraud detection scenarios testing environments featuring up to 40% malicious nodes maintained normal classification accuracy rates at around 89%, all while controlling communication loads below traditional approaches by over 70%.
Systematic Research Advances in Vertical Federated Learning
n **Technical Framework for Cross-industry Data Collaboration** n Springer’s review paper published on vertical federate learning (Vertical FL,VFL) establishes complete technical maps within this domain for the first time compared against horizontal federate learnings’ unique challenge concerning aligning heterogeneous feature spaces held by differing organizations—for instance “bank-e-commerce” joint risk control models wherein banks possess user credit records whilst e-commerce platforms hold consumer behavior data showcasing both dimension disparities alongside notable statistical distribution shifts.” The paper thoroughly compares seven mainstream feature alignment methodologies highlighting pros/cons respectively: n - Encryption-based entity resolution schemes exhibit excellent privacy but increase computational complexity quadratically relative sample sizes; n - While utilizing federal embedding mapping achieves higher efficiencies necessitating predefined public anchor point datasets.” On matters pertaining privacy protections trends towards hybrid usage homomorphic encryption(HE) alongside secure multiparty computation(SMPC)—e.g., during clinical trial data-sharing instances whereby HE handles numerical examination indicators whereas SMPC oversees collaborative calculations diagnostic results." Authors emphasize future VFL developments require establishing standardized cross-industry interfaces including foundational components such metadata description languages unified modeling protocols audit tracking mechanisms etc."\
Privacy-Efficiency Balancing Practices Within Medical Contexts \
**Optimizing Differential Privacy Applications Cancer Diagnosis" Published Nature Scientific Reports'25 Clinical Study sets new benchmarks applications sensitive healthcare datasets collaboration seventeen oncology specialty hospitals constructing largest distributed breast cancer imaging analytics platform thus far contributing significantly proposing dynamic privacy budget allocation algorithms able adjust Gaussian noise injection intensity according quality ratings assigned annotations provided respective healthcare providers . Specifically ,relaxed budgets allocated facilities exhibiting high consistency ε=2 .1 strict protections applied contentious cases ε=1 .3 Technical implementations leverage layered protective frameworks pixel-level differential processing original DICOM images layer-wise applying intermediate deep networks final stage imposing gradient noise additionally achieving robust defense measures ensuring resulting models attain accuracies reaching96 .1 % merely0 .9 percentage points lower centralized training fully compliant HIPAA mandates surrounding untraceable individual data requirements also pioneering clinical interpretability standards demanding predictions accompanied consistent reports validated inter-institutionally confirming overall integrity outcomes obtained throughout trials conducted herewith presenting innovative solutions fostering advancements beyond current limitations established prior iterations deployed accordingly.