Federated Learning is a distributed machine-learning approach in which multiple participating devices or organisations collaboratively train a model while keeping their local training data on the local participants rather than centrally pooling all raw data. A typical federated-learning process sends an initial model to participating clients, performs local training on each client's data, sends model updates or related information to a coordinating server, aggregates contributions into a new global model and repeats the process.
The raw datasets therefore remain at their originating locations under the standard formulation. Federated learning was developed to address settings where centralising data is undesirable or impractical because of privacy requirements, data ownership, regulatory constraints, network costs or organisational boundaries. It is not itself a guarantee of privacy: model updates can potentially reveal information, and additional protections such as secure aggregation, differential privacy or encryption may be required depending on the threat model.
Federated learning can operate in cross-device and cross-silo settings. Cross-device federated learning involves potentially very large numbers of consumer or edge devices, often with limited compute, intermittent connectivity and heterogeneous hardware. Cross-silo federated learning generally involves a smaller number of relatively reliable organisations, such as hospitals, banks or enterprises, that hold substantial local datasets. These environments create different engineering and statistical challenges.
Client availability, communication bandwidth, battery consumption, storage capacity and local computation can affect cross-device training, while organisational governance, data heterogeneity and coordination can be more prominent in cross-silo systems. The participating datasets are often non-identically distributed, meaning different clients may have different populations, class frequencies, feature distributions or local environments. This non-IID structure can make straightforward aggregation less effective than it would be under centrally pooled, identically distributed data.
Federated averaging is a foundational aggregation method in which clients perform local optimisation and a server computes a weighted average of their updated model parameters, often weighting contributions according to local dataset size. Local training reduces the frequency with which raw examples or individual gradients must be exchanged, but excessive local training can increase divergence between client models when their datasets differ substantially. Other federated optimisation methods modify the aggregation or local update procedure to improve convergence under heterogeneous conditions.
Client sampling is often used because not every participant can or should communicate during every training round. Systems may also use asynchronous participation, personalised models or clustered approaches when a single global model is inappropriate. The optimisation problem is consequently influenced by both machine-learning considerations and distributed-systems constraints.
Privacy and security are central to the motivation for federated learning, but they require careful threat modelling. Keeping raw data on a device reduces the need to transmit the original dataset to a central server, but model updates can still contain information about local training examples. An adversary may attempt to infer sensitive information from updates, reconstruct aspects of training data or introduce malicious updates.
Secure aggregation can allow a server to obtain an aggregate update without directly seeing each participant's individual contribution. Differential privacy can add formal bounds on information leakage by introducing carefully calibrated randomness, though it generally introduces a privacy-utility trade-off. Encryption and trusted execution environments can provide additional protections in some architectures. Federated learning should therefore be understood as a data-localisation strategy that can form part of a broader privacy and security architecture, not as a synonym for private machine learning.
Federated learning is useful when organisations possess complementary datasets but cannot simply combine them. Several institutions may wish to train a shared model while retaining control over their local records. A network of devices can also contribute usage information without routinely uploading all raw data to a central repository. Potential application areas include mobile personalisation, healthcare research, financial modelling, industrial monitoring and edge-device intelligence.
In healthcare, for example, different institutions may have locally governed datasets whose direct centralisation presents legal, ethical or operational difficulties. Federated training can potentially allow models to benefit from patterns across institutions, although differences in equipment, patient populations, coding practices and measurement protocols create substantial statistical challenges. The value of federated learning therefore depends on both the ability to train effectively across heterogeneous datasets and the governance arrangements governing participation, model use and information sharing.
Communication is a major systems bottleneck in federated learning. A model may contain millions or billions of parameters, and repeatedly transmitting complete parameter sets or updates can be expensive. Compression, sparsification, quantisation and other communication-efficient methods can reduce bandwidth requirements. Client devices may also have limited processing power, memory or battery capacity, making the amount of local computation an important design parameter. Training must tolerate unreliable clients that disconnect, fail or provide updates at different speeds.
Security threats include poisoned updates, sybil participation, compromised devices and malicious coordinators or participants. Robust aggregation techniques can reduce the influence of abnormal updates, but no single aggregation method eliminates all attacks. Federated learning is consequently a combined problem in optimisation, distributed systems, privacy, security and statistical heterogeneity.
Federated learning differs from centralised distributed machine learning because the raw training data remain distributed across participants rather than being collected into a common central dataset. It also differs from decentralised learning architectures in which there may be no central coordinator. Federated learning can use supervised, unsupervised, semi-supervised or self-supervised objectives depending on what local data and targets are available. Deep neural networks are commonly used but are not a defining requirement.
Transfer learning can be incorporated by sending a pretrained model to clients and adapting it locally. Reinforcement-learning components can potentially participate in federated schemes when appropriate distributed optimisation objectives exist. The important distinction is therefore about where data and training computation reside and how model information is exchanged, not about whether a particular neural-network architecture is used.
The advantages of Federated Learning include reduced need for raw-data centralisation, potential improvements in data governance, the ability to learn from geographically or organisationally distributed datasets and the possibility of adapting models to local environments. Its limitations include communication overhead, heterogeneous client data, unreliable participation, complex security requirements, additional system engineering and the possibility that local distributions are too different for one global model to perform well.
Privacy benefits depend on implementation and threat model, and aggregation does not automatically prevent inference or poisoning attacks. Federated systems also require mechanisms for participant authentication, update validation, model versioning, auditing and potentially compensation or incentive structures. When local data are highly heterogeneous, personalised federated learning may be more appropriate than a single global model. Thus, federated learning changes the data architecture of machine learning but does not remove the fundamental requirements for reliable modelling, evaluation, security and governance.
