What Is Federated AI?

Businesses in all different sectors are excited about the opportunities that AI presents. But they also recognize there are challenges they must overcome to capitalize on those opportunities.

Many of the biggest challenges involve creating an AI-ready data strategy. These include:

• Keeping sensitive data protected and within prescribed borders to address the complexities of the current geopolitical landscape
• Finding enough data to train accurate, comprehensive models
• Sharing AI model weights with partners without sharing the underlying raw data
• Processing massive volumes of data quickly and efficiently
• Moving data between distributed processing locations without delay

Overcoming all these challenges at once certainly won’t be easy. Fortunately, there are steps enterprises can take to help simplify things. Adopting federated AI is chief among them.

Federated AI vs. centralized AI

In traditional AI model training, organizations move datasets from many distributed sources to a single processing location. This practice is known as “centralized AI.” Federated AI, also known as federated learning, has emerged as a valuable alternative.

Instead of training a global model all in one place, organizations train smaller models locally at various edge sites. This removes the need to transfer massive amounts of raw data to the processing location. Instead, they can transfer only the model weights—the numerical parameters that reflect what the model learned during the training process. Organizations can aggregate different model weights to form a single global model. Finally, they’ll push that global model back to the edge locations where they’ll perform inference.

Features of centralized AI vs. features of federated AI

What are the different varieties of federated AI?

There are two main approaches to federated AI:

• Horizontal federated AI, which pulls model weights from the same types of data in every site
• Vertical federated AI, which pulls model weights from different types of data in different sites

Most organizations will pursue a combination of vertical and horizontal federated AI, based on the availability of different datasets. Vertical federated AI comes with unique challenges, such as the need to perform entity matching to safeguard against data leakage[1], but the benefit is typically worth the effort.

Also, federated AI can take place within a single organization or across different organizations. In the latter case, federated AI allows business partners to share their model weights and collaborate to form a global model. This helps them benefit from each other’s datasets without placing the underlying data at risk. It also shares the energy burden of training the model, which can help deploy AI without derailing sustainability goals.

When different organizations do federated AI together, they’ll typically create an AI marketplace overlay to enable a more systematic approach to data sharing. This allows them to monetize their AI datasets and establish governance principles for how data gets shared within the marketplace.

Organizations can apply a variety of privacy-enhancing techniques to prevent raw data from being exposed to AI processing. These include:

• Differential privacy, a technique that involves injecting statistical noise into datasets to better protect the raw data
• Secure multiparty computation, which allows all participating parties to see the outcome of a computation without seeing the inputs provided by the other parties

Why are organizations pursuing federated AI?

Organizations see federated AI as a way to build out their AI-ready data strategy while also protecting their sensitive data. This is because federated AI allows them to share AI model weights without the raw data leaving their security perimeter. When they’re able to maintain control over their AI datasets, they’re better positioned to meet their data privacy and sovereignty requirements.

Federated AI also allows them to significantly reduce the amount of data they transfer. In fact, one group of Equinix customers reduced their data transfer burden by more than 99.9% compared to a centralized training model. This is important because moving very large datasets contributes to higher costs, lower performance and decreased energy efficiency.

Finally, many organizations are pursuing federated AI because they recognize that training a precise model requires a lot of data, and they may have difficulty getting enough data internally. Federated AI provides a way for all the partners in an ecosystem to contribute to a global model that benefits everyone, without having to sacrifice data privacy.

How are organizations using federated AI?

Federated AI is helping enterprises pursue a variety of different industry use cases, including:

• Anti-money laundering: Banks can perform federated AI internally and share the results with partners, service providers and regulators. This can help detect patterns that suggest money laundering activity.
• Cybersecurity threat detection: Coalition partners, like those that make up NATO, can share threat insights in real time. This allows the partners to track potential threats as they develop.
• Cancer research: Hospitals and research organizations can share insights with one another to accelerate research and improve patient outcomes. Crucially, they can do this without exposing data from individual patients.
• Autonomous vehicles: Different providers and manufacturers can share traffic safety insights. This leads to vehicles that are better prepared to identify and avoid obstacles in real time.
• Airplane maintenance: Airlines can share sensor data about individual components within their airplanes. This allows them to aggregate enough data to create accurate models for predictive maintenance.

What are the challenges of federated AI?

Federated AI requires data consistency across all participants. Some participants may sample their local datasets differently. This can lead to model weights that aren’t representative, which in turn introduces bias into the global model.

Ensuring data consistency is just one example of why governance is so important for federated AI. A strong governance model should define how partners share data, who can use the insights from that data, how each partner can check the quality of the data from other partners, and how partners can stop sharing data when they no longer wish to.

There’s also the fact that federated AI is more complex to deploy than centralized AI is. Click-and-deploy AI factory solutions are becoming the norm for centralized AI. It’s true that there are similar solutions intended to make federated AI easier to deploy, including TensorFlow Federated, Flower, KubeFATE and PySyft. But for the most part, federated AI still requires specialized expertise to make sure all the different partners and data sources are connected correctly.

Finally, federated AI isn’t a good fit for every situation, and not every AI/ML algorithm can be operated in a federated manner. There will always be a tradeoff between data transfer requirements and model accuracy. It’s true that federated AI provides lower model accuracy, but for many use cases, that accuracy is still well within an acceptable range. For others, it isn’t. Organizations must carefully evaluate each individual use case before deciding whether to proceed with federated AI.

Where should organizations do federated AI?

Organizations looking to get started with federated AI will need digital infrastructure in different locations, as well as dedicated networking solutions to connect those locations.

With a global platform of data centers in 74 markets across 35 countries, Equinix is uniquely suited to help our customers pursue federated AI. Many of those markets are connected by Equinix Fabric®, our virtual interconnection solution. This helps customers move data and models between sources and aggregation points, quickly and securely.

For global enterprises that can’t move certain datasets outside of certain borders, Equinix’s global reach helps them deploy AI nodes in all the right places to meet their data sovereignty requirements. The example map below shows how a customer could deploy federated AI clusters in three different regions, each connected to the other two by Equinix Fabric.

In addition, Equinix enables lower latency for AI inference workloads. Most major metro areas are within 10 milliseconds round-trip time (RTT) of an Equinix IBX® data center. This means that our customers can host AI workloads closer to the data source, removing the need for costly, time-consuming data backhauls.

For customers looking to do federated AI with partners, Equinix provides a neutral platform on which to host the model aggregation point. Each individual partner can connect securely to the shared infrastructure from wherever they’re based.

Learn more about how data privacy and compliance fit into an AI-ready infrastructure strategy: Read the IDC analyst brief Private AI Infrastructure Solves for Privacy and Regulatory Compliance Requirements.[2]

[1] Yang Liu et al, Vertical Federated Learning: Concepts, Advances and Challenges, IEEE Transactions and Data Engineering, 2024.

[2] Dave McCarthy, Private AI Infrastructure Solves for Privacy and Regulatory Compliance Requirements, IDC analyst brief sponsored by Equinix, April 2024.