Federated Learning: A decentralized approach to smarter, safer AI

Authored by: Raquel Lazcano and Maria Alejandra Paz (ATOS) 

Federated Learning represents a transformative shift in how Artificial Intelligence (AI) models are developed and deployed. Unlike traditional machine learning, where data must be centralized in one place for training, Federated Learning enables model training to occur directly across multiple distributed devices or nodes. This approach makes it possible to use the collective intelligence of diverse data sources, without compromising privacy or data ownership. 

Federated Learning can be applied to multiple sectors where data is distributed, privacy is critical, or connectivity is limited, like in healthcare, IoT, edge computing, and manufacturing, among others. 

At its core, Federated Learning is characterized by: 

• Decentralization, distributing model training across devices or nodes  

• Privacy preservation, ensuring sensitive data remains on-device 

• Reduced data transfer overhead, minimizing the need for extensive data movement 

Federated Learning is especially useful in contexts where: 

1. Data volumes are too large to transfer efficiently, making centralized training impractical or expensive. 

1. Data is highly sensitive, governed by privacy regulations or intellectual property restrictions that prevent it from leaving local environments. 

1. Collaborative learning across complementary datasets is desired, where each participant holds different features about the same entities. 

Key benefits and strengths 

Federated Learning introduces a new way to collaboratively train AI models while respecting data privacy and optimizing resources. Its benefits extend from communication efficiency to industrial-scale validation, making it an effective and trustworthy approach for distributed AI development. 

1. Privacyand security by design 

• Data never leaves the client’s premises, maintaining anonymity and ensuring compliance with privacy regulations. 

• Communication between nodes is encrypted, and privacy-preserving protocols can be applied for additional protection. 

• The decentralized and distributed structure minimizes exposure risks and removes the need for a central data repository. 

2. Efficiencyand scalability 

• Only model updates (weights or gradients) are exchanged instead of raw data, significantly reducing communication and network overhead. 

• Compression techniques further streamline communication and lessen infrastructure strain. 

• Each participant manages its own storage and computational resources, allowing flexible scaling based on local capabilities. 

3. Collaborationon model quality 

• Multiple organizations can jointly train a shared model without sharing data, leading to richer, more generalizable, and higher-quality AI models. 

• The modular and flexible design supports a wide range of use cases and easily integrates diverse privacy-preserving algorithms. 

4. Continuallearning 

• Integration with MLOps tools enables continuous monitoring, automatic retraining, and adaptation to new data or changing conditions. 

• Proven performance in real-world industrial pilots demonstrates the robustness and reliability of the approach. 

Implementation  

In ALCHIMIA, Federated Learning is a core technology to enable secure, privacy-preserving, and collaborative AI across industrial partners. Specifically, one of the main technical outcomes of ALCHIMIA is Atos FL framework, which allows defining computational graphs to implement Federated Learning tasks. It is based on a pipes and filters design pattern to model the federated actors, the operations they perform and the exchange of information between them. Its main features are: 

1. Modularity, interoperability and customizability: 

The framework, provided as a Python library, is based on a modular architecture based on a set of functional units that can be inherited and extended, allowing new components to be implemented or custom functions to be injected to fully personalize the federated pipeline. This interoperability and customizability are clearly showcased in the native support to the main ML frameworks, such as Keras, PyTorch, and Scikit-learn, or in the seamless integration with diverse communication protocols, such as HTTPS, Kafka, MQTT and gRPC, among many other features. 

1. Strong privacy-preserving and secure aggregation mechanisms: 

The framework has a strong focus on privacy preservation. To that end, the framework supports several privacy-preserving mechanisms, such as Gaussian noise addition, secure-sum aggregation, and TLS-based channel encryption. To complement this, the framework also supports advanced secure aggregation algorithms, including FedAvg, FedOpt variants, Krum, Median, Trimmed-Mean, GeoMedian, and Tree Bagging, supporting both robust aggregation (resistant to outliers and adversarial updates) and optimization-oriented strategies. 

1. Flexibility and adaptability: 

The modularity and adaptability of the framework also makes it very flexible, allowing the implementation of complex topologies, such as swarm learning or hierarchical learning. 

1. Efficiency and scalability:  

The framework is also being developed with efficiency in mind, integrating model compression capabilities including both lossless (zstd, lzma) and lossy (quantization via QSGD) techniques to reduce communication overhead. 

1. Third-party integration:  

One of the framework’s key features is its compatibility with mainstream ML tools and MLOps frameworks, such as MLFlow, MinIO, Grafana, Prometheus and Evidently to support experiment tracking, artifact storage, performance monitoring, data drift detection, and continuous model evaluation workflows. This allows organizations to embed FL into their AI pipelines, simplifying deployment, monitoring, and continuous improvement.