Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence
Edward Collins, Michel Wang · 2025
Read Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence by Edward Collins, Michel Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-17703.txt.
Introduction
Ratun Rahman surveys federated learning, a way for many devices or institutions to train one shared model without pooling their private data, in a survey paper from 2025. Phones, hospitals, and banks hold valuable data they cannot legally or safely centralize. Federated training keeps data local and shares only model updates.
The survey walks through the standard cycle of local training, merging, and global updates. It weighs the field's main hurdles: mismatched data across clients, varied hardware, heavy communication costs, and privacy protection. Trends like personalization and quantum links get coverage too. Benchmarks and real uses round out the guide.
Catalog record
| CREATOR | Edward Collins, Michel Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-17703
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence Edward Collins, Michel Wang Abstract Federated Learning (FL) has emerged as a transformative paradigm in the field of distributed machine learning, enabling multiple clients—such as mobile devices, edge nodes, or organizations—to collaboratively train a shared global model without the need to centralize sensitive data. This decentralized approach addresses growing concerns around data privacy, security, and regulatory compliance, making it particularly attractive in domains such as healthcare, finance, and smart IoT systems. This survey provides a concise yet comprehensive overview of Federated Learning, beginning with its core architecture and communication protocol. We discuss the standard FL lifecycle, including local training, model aggregation, and global updates. A particular emphasis is placed on key technical challenges such as handling non-IID (non-independent and identically distributed) data, mitigating system and hardware heterogeneity, reducing communication overhead, and ensuring privacy through mechanisms like differential privacy and secure aggregation. Furthermore, we examine emerging trends in FL research, including personalized FL, cross-device versus cross-silo settings, and integration with other paradigms such as reinforcement learning and quantum computing. We also highlight real-world applications and summarize benchmark datasets and evaluation metrics commonly used in FL research. Finally, we outline open research problems and future directions to guide the development of scalable, efficient, and trustworthy FL systems. I Introduction In recent years, the landscape of machine learning (ML) has been rapidly transformed by the explosive growth in data generation from edge devices such as smartphones, wearables, autonomous vehicles, and smart sensors. These devices collectively generate petabytes of data daily, representing a vast and valuable source of information for training intelligent systems [ 1 , 2 ] . However, centralizing such data in cloud servers for model training has become increasingly impractical due to multiple limitations—chief among them being privacy concerns, communication overhead, and compliance with data protection regulations such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA) [ 3 ] . Federated Learning (FL), proposed by McMahan et al. [ 4 ] , addresses these limitations by introducing a collaborative yet privacy-preserving machine learning framework. In FL, training is conducted locally on edge devices or institutional servers, and only model updates (e.g., weights or gradients) are transmitted to a central aggregator. This paradigm allows raw data to remain decentralized, ensuring both privacy preservation and reduced bandwidth consumption. A central server then performs aggregation, most commonly using Federated Averaging (FedAvg), to update the global model iteratively [ 4 ] . FL has quickly evolved from its original scope—primarily involving mobile devices in cross-device scenarios—to more complex, structured environments such as cross-silo learning, where institutions like hospitals, banks, or universities collaboratively train models on sensitive datasets [ 5 ] . These two categories present vastly different challenges. Cross-device FL must contend with large-scale, unreliable, and computationally limited clients, while cross-silo FL emphasizes security, trust, and statistical heterogeneity among fewer but more powerful clients [ 2 ] . One of the most pressing technical challenges in FL is the presence of statistical heterogeneity or non-IID (non-independent and identically distributed) data across clients. Since each client’s data may be generated under different distributions—due to geographic, demographic, or temporal variations—traditional centralized learning assumptions often do not hold in federated settings [ 1 , 6 ] . This heterogeneity can significantly degrade convergence speed and model accuracy, necessitating robust aggregation strategies, personalized learning techniques, and client clustering mechanisms [ 7 , 8 ] . Moreover, FL must cope with system-level heterogeneity, where client devices vary in terms of computing power, memory, battery life, and network connectivity [ 9 ] . Solutions such as asynchronous training, partial participation, and client selection heuristics have been proposed to alleviate these constraints [ 10 , 11 ] . Communication efficiency is another core issue, as iterative communication rounds can be costly and slow in real-world deployments. Techniques such as update compression, quantization, and sparsification have emerged to reduce the size of model updates and improve communication throughput [ 12 ] . Privacy remains central to the FL paradigm. While FL inherently reduces data exposure, model updates may still leak sensitive information through gradient inversion attacks or malicious aggregation [ 13 ] . To counter such risks, researchers have developed secure aggregation protocols [ 14 ] and incorporated differential privacy mechanisms at the client level [ 15 ] . These methods offer mathematical guarantees of privacy, albeit at the cost of reduced model accuracy or increased computation. The field of FL continues to grow, integrating innovations from other domains such as reinforcement learning [ 16 ] , multi-task learning [ 7 ] , blockchain [ 17 ] , and even quantum computing [ 18 ] . These interdisciplinary expansions aim to enhance FL’s scalability, robustness, and privacy assurances. Applications of FL have already made impactful inroads in healthcare (e.g., disease diagnosis from medical imaging), finance (e.g., fraud detection), smart cities (e.g., traffic prediction), and natural language processing (e.g., next-word prediction in mobile keyboards) [ 19 , 20 ] . This survey offers a concise yet comprehensive overview of Federated Learning. We begin by outlining its foundational concepts and architectural designs, followed by an analysis of its key challenges, including statistical and system heterogeneity, communication bottlenecks, and privacy threats. We then review recent innovations and practical implementations across various domains. Finally, we discuss ongoing research trends and open problems, laying the groundwork for future exploration and advancement in this promising field. II System Architecture Federated Learning (FL) operates under a client-server architecture, where a global model is collaboratively trained by multiple clients without exchanging their raw data. This architecture is designed to preserve data privacy, reduce communication overhead, and enable scalability across diverse hardware environments [ 5 , 2 , 21 ] . In this section, we describe the core architectural components of a typical FL system, focusing on both centralized and decentralized settings. II-A Centralized Federated Learning In the standard FL setting introduced by McMahan et al. [ 4 ] , a central coordinator (or server) orchestrates the training process. The architecture consists of three main stages: (1) the server distributes the current global model to selected clients; (2) clients perform local training using their private data; and (3) the server aggregates the clients’ updates to form a new global model. This process is repeated iteratively until convergence. The most commonly used aggregation algorithm is Federated Averaging (FedAvg), where each client trains the model for a few epochs and sends updated parameters to the server. The server then performs a weighted average of the received updates based on the clients’ local data sizes [ 4 ] . II-B Decentralized and Peer-to-Peer FL To mitigate the single point of failure and trust issues associated with centralized architectures, decentralized FL frameworks have emerged. In such systems, clients coordinate among themselves without relying on a central server [ 22 , 23 , 24 ] . Techniques such as gossip-based communication, ring topologies, and blockchain are used to enable secure and robust peer-to-peer model sharing. For example, blockchain-integrated FL ensures tamper-proof logging of updates and consensus-driven aggregation without a centralized entity [ 17 ] . While these methods increase resilience and transparency, they often incur additional latency and computational complexity. II-C