Federated learning: Decentralised training for privacy-preserving AI

Download

Federated learning allows multiple entities to collaboratively train and fine-tune AI models in a decentralised manner, ensuring data privacy by eliminating the need to share raw data. This article explores how federated learning works, its key applications, and the industries poised to benefit most from this approach.

What is federated learning?

Federated learning is an approach that allows multiple distributed entities (e.g. devices, servers, organisations) to collectively train and finetune an AI model without sharing any raw data. This differs from traditional approaches that centralise data to use for training of the model.

While consumer use cases for federated learning exist, this article focusses on the enterprise space.

Defining the types of federated learning

As this is a relatively new approach, there are different, sometimes contradicting, definitions of federated learning. We at STL define types of federated learning by the location of data processing for training, and the parties involved in the training (summarised in Figure 1).

Figure 1: STL’s definitions of different types of federated learning

Source: STL Partners

How it works?

Each entity (e.g. for device, server, organisation) involved in training/ finetuning maintains its own copy of the model. Training occurs locally using raw data that never leaves the entity.

For example, in a predictive maintenance use case in manufacturing, IoT sensors collect data that is stored and processed on an on-premise server. The server trains the model by updating its parameters based on patterns in the raw data. In this case, while raw data moves from the IoT sensor to the server, it remains within the same entity (the server) and is never shared externally.

Each entity independently trains its model using its own raw data, generating a set of local parameter updates. Instead of sharing raw data, these updates are sent to a central server, which aggregates updates from all entities to compute an optimal global update to the model parameters. The updated parameters are then distributed back to all entities.

This approach enables collaborative model training across multiple organisations or within a single organisation. For instance, in the automotive industry, vehicle data collected in one country may not be shareable with a business unit in another due to regulatory or data sovereignty constraints. This method allows organisations to benefit from collective learning while ensuring compliance with data privacy regulations.

What are the main benefits of federated learning?

Collaboratively training AI models offers clear advantages. By leveraging diverse datasets, models improve in accuracy, generate more useful insights, and produce better recommendations. This diversity is particularly valuable for generative AI, enabling it to create more unique and innovative outputs.

While these benefits can be achieved through traditional centralised training, where all data is aggregated in one location, federated learning has two unique advantages: data privacy and reduced backhaul (illustrated in the diagram below).

Of these, data privacy is the most significant, alone it can be enough of a driver to justify investment into federated learning for specific use cases. Reduced backhaul is a valuable secondary benefit, but alternative solutions, such as edge-based pre-processing, exist. In contrast, ensuring data privacy while collaboratively training models is significantly harder to achieve through conventional methods.

Figure 2: Main benefits of federated learning compared to centralised model training alternative

 

Source: STL Partners

Which applications and verticals will benefit most from federated learning?

Banking and finance

Examples of organisations that may benefit from collaboratively training AI models: Banks, financial institutions, and payment processors

Fraud prevention

Types of data use to train AI model: Transaction data across multiple banks and payment networks

Enhances fraud detection by improving the ability of AI models to identify suspicious behaviour, such as unusual spending patterns or device inconsistencies, by comparing them to broader fraud trends across a diverse range of institutions.

Figure 3: Multiple banks collaboratively training an anti-fraud AI application in a cross-silo consortia model

Anti-money laundry

Types of data use to train AI model: Transaction data across multiple banks and payment networks

Helps individual banks detect money laundering by identifying layering transactions, structured fund transfers, and unusual account activity that align with known laundering techniques.

Healthcare

Examples of organisations that may benefit from collaboratively training AI models: Hospitals and clinics, research institutions, pharmaceutical companies, and public health agencies

Predicting public health trends

Types of data used to train AI models: Electronic health records, epidemiology and public health data, and environmental health data from multiple healthcare institutions

Improves disease prediction by aggregating information about the health of a local area from multiple data sources and allowing that to be compared with broader epidemiological trends.

Pharmaceutical

Drug discovery and development

Types of data used to train AI models: Molecular structures, genetic data, and clinical trial results from pharmaceutical companies and research institutions

Enables pharmaceutical companies to screen their own compound libraries more effectively, identifying potential treatments that match patterns discovered across the broader research community.

Manufacturing

Examples of organisations that may benefit from collaboratively training AI models: Manufacturers, suppliers, and research teams to collaborate across facilities, companies, and geographies

Product design and development

Types of data used to train AI models: Customer feedback, industry trends, and internal engineering data from multiple teams and production sites

Identifies recurring design flaws, optimises material selection, and improves durability. For example, aerospace companies can analyse sensor data from aircraft fleets to detect structural fatigue patterns, enhancing maintenance strategies and safety.

Autonomous vehicle

Types of data used to train AI models: Sensor data from multiple vehicle fleets, including camera feeds, LiDAR scans, and real-time traffic conditions

Improves onboard AI models for autonomous driving, enhancing obstacle detection and adaptive decision-making while keeping raw driving data local and private to each manufacturer or operator.

Conclusion

Federated learning has the potential to change how AI models are trained, enabling access to more diverse and representative data without compromising privacy. By shifting away from traditional centralised training, it allows organisations to build more accurate, robust, and generalisable AI models. While challenges around scalability, standardisation, and security remain, continued advancements in infrastructure and governance could establish federated learning as a key enabler of AI in highly regulated and data-sensitive industries.