Federated learning is a machine-learning approach in which multiple participating devices or organisations collaboratively train a model while keeping their local training data at or near its original location instead of centrally pooling all raw data. A central coordinator, federated server or peer-to-peer mechanism distributes a model or model parameters to participants; participants train locally using their own data; and they return selected model updates or other training information for aggregation.
The aggregated model is then redistributed and the process repeats. The central objective is to enable collaborative learning across distributed datasets while reducing the need to transfer raw training data to a central repository. Federated learning is therefore primarily a distributed optimisation and privacy-oriented training architecture, not a guarantee that data is completely private. Model updates can potentially reveal information, and federated systems require additional security and privacy techniques when strong confidentiality is required.
A standard federated-learning procedure is often described using a cycle of model distribution, local computation, update transmission and aggregation. In a common approach known as Federated Averaging, a server sends the current global model to a selected group of clients. Each client trains that model for some number of local steps or epochs on its own data and calculates updated parameters.
The server then aggregates the client updates, often weighting them according to the amount of training data held by each participant, to produce a new global model. This process can repeat over many communication rounds. The raw local examples need not be sent to the server. However, the server still receives information derived from those examples through updates, and the security of the communication and aggregation process therefore remains important. Variations can change the aggregation rule, client selection, local objectives, communication frequency or model architecture.
Federated learning is especially relevant when data is distributed for practical, legal, commercial or privacy reasons. Mobile devices can generate local interaction data that may be undesirable or impractical to upload centrally. Hospitals or financial institutions may hold sensitive datasets that cannot easily be pooled. Industrial organisations may have proprietary operational data that they are unwilling to transfer to another party. Federated learning can allow each participant to contribute to a shared model while retaining control of its raw records. However, technical feasibility does not automatically imply legal permission or privacy compliance, and federated learning does not remove governance requirements. Organisations still need to consider consent, data-processing obligations, security, retention, access controls, model governance and the possibility that derived information can expose sensitive properties.
Federated learning introduces several forms of heterogeneity. In cross-device federated learning, potentially millions of devices such as phones may participate intermittently, often with limited battery, bandwidth and computational capacity. In cross-silo federated learning, a smaller number of relatively powerful organisations such as hospitals, banks or companies collaborate, usually with more stable connectivity and stronger administrative control. Data across participants can be non-independent and identically distributed: one device or organisation may contain a substantially different population or behaviour pattern from another.
This non-IID distribution can make naïve averaging less effective because local updates may point in conflicting directions. Client availability can also vary, and some participants may contribute far more examples than others. Federated algorithms therefore often incorporate strategies for client selection, weighting, local training, communication efficiency and robustness to heterogeneous data.
Privacy and security mechanisms are important because “data never leaves the device” is not equivalent to “the training process reveals nothing”. A malicious or curious participant may attempt to infer information from model updates, while an attacker could manipulate updates to poison the global model. Secure aggregation protocols can allow a server to obtain an aggregate of client contributions without directly seeing each individual update.
Differential privacy can add carefully calibrated noise to reduce the ability to infer whether particular data contributed to training, although privacy comes with utility and accuracy trade-offs. Encryption, authentication, access controls, anomaly detection and robust aggregation can address other threats. Federated learning can therefore be part of a privacy-preserving architecture, but the actual privacy guarantees depend on the complete system design rather than on federated training alone.
Federated learning differs from conventional distributed machine learning in an important architectural respect. In ordinary distributed training, multiple compute workers may access a centrally stored dataset or centrally controlled partitions of the same dataset, with the primary goal often being faster optimisation. Federated learning instead assumes that data remains decentralised and that participants may have different ownership, privacy and operational constraints. The distinction is not absolute because both involve distributed optimisation, and some systems use hybrid designs. Federated learning also differs from decentralised learning in which there is no central coordinating server; decentralised approaches can exchange updates directly among participants. The term therefore identifies a broad family of collaborative training architectures rather than one single algorithm.
Federated learning has practical limitations. Communication can be expensive because models or updates may contain millions or billions of parameters. Participants may disconnect, use heterogeneous hardware or train for different amounts of time. Non-IID data can slow convergence or produce a global model that performs poorly for minority participants. Malicious clients can conduct poisoning or model-update attacks, while a compromised server can threaten confidentiality.
Fairness is also a technical and governance concern: a global model optimised for aggregate performance may underperform for participants whose data distributions are rare or poorly represented. Personalisation techniques can address some of these issues by adapting a shared model to individual participants. Compression, quantisation, sparsification and local-update strategies can reduce communication requirements, but they can alter optimisation behaviour or model accuracy.
Federated learning is therefore best understood as collaborative model training over decentralised data, with local data retained by participating clients and model information exchanged instead. Its principal motivations are data locality, privacy, governance, communication constraints and the practical impossibility or undesirability of pooling all raw data. Its principal challenges include statistical heterogeneity, communication costs, device availability, malicious participants, update leakage and the difficulty of providing strong privacy guarantees.
It is particularly suitable where multiple parties can agree on a common modelling objective but cannot or should not centralise their underlying datasets. In production systems, federated learning normally needs to be combined with secure communication, authentication, secure aggregation, privacy mechanisms, monitoring and governance. The technique therefore changes where computation and data reside, but it does not by itself eliminate the security, privacy or statistical risks associated with machine learning.
