Fighting Crime Without Compromising Privacy: Federated Learning in PRESERVE

By PRESERVE partner Bournemouth University

Why Federated Learning matters

Today we’re diving into the fascinating world where cutting-edge artificial intelligence meets the real-world demands of privacy, ethics, and cross-border collaboration. If you’ve ever wondered how organisations like hospitals, banks, and police forces can work together to train smarter AI—without putting sensitive data at risk—you’re in the right place.

Why sensitive data is both a blessing and a burden

Let’s face it: the most useful data for training AI is often the most sensitive. Think medical records, financial transactions, or police case files. These are treasure troves for machine learning but moving them around or centralising them is a privacy (and legal) nightmare. With regulations like GDPR guarding personal information, sharing such data is unlawful and outright forbidden.

Turning the data game upside down

Federated Learning (FL) is a technology that flips the script. Instead of shipping all the data to a central AI model, FL sends the model to where the data lives. Each participant, be it a hospital, a police agency, or even your smartphone, trains the AI locally and only shares the lessons learned, but never raw data, with the peers. This keeps sensitive data locked down while still letting everyone benefit.

Because sensitive data stays where it originated, FL significantly enhances privacy and reduces the susceptibility to data breaches and misuse that are often associated with centralised data handling.

Figure 1 compares classical distributed learning with the two main flavours of federated learning: centralised and decentralised FL. In classical distributed learning, data is transferred to a central location for model training. By contrast, federated learning keeps data with each participant. In centralised FL, a coordinating server orchestrates the process: it collects the participants’ model updates, aggregates them, and redistributes the improved global model. In decentralised FL, the central server is removed entirely, and participants exchange model updates directly with one another through a peer-to-peer network, which can improve resilience and remove reliance on a central coordinator but increases communication complexity. Whatever the flavour, every FL system rests on three core components: the clients, which hold local data and independently compute model updates; a communication framework, which transports those updates; and an aggregation algorithm, which merges the incoming contributions into a shared model so that everyone benefits from the diversity of local datasets without compromising the privacy of the underlying data.

Figure 1: Comparison of (a) classical distributed learning, (b) centralised federated learning, and (c) decentralised federated learning.

In its typical centralised configuration (Figure 2), FL proceeds through repeating cycles, called rounds, of two-way communication between the coordinating server and the participating clients.

Share the starting model. Every client begins by downloading the most recent version of the shared global model from the server, which becomes the starting point for its local model.
Learn locally. Each client trains that model on its own dataset, right where the data lives. Nothing leaves the building — only the local copy of the model improves.
Send back the lessons. Once local training is complete, each client returns only its improved model — the small adjustments it has learned — to the server. The underlying data is never transmitted.
Combine the knowledge. The server aggregates all incoming local models into an enhanced global model. The classic method, Federated Averaging (FedAvg), computes a weighted average of the clients’ model parameters, giving more weight to participants that trained on larger datasets.
Repeat. The upgraded global model is sent out again and the cycle continues until the model reaches the desired accuracy, or the server decides to end the training.

This simple loop has powerful consequences. With every round, each participant benefits from what all the others have learned, without any of them ever seeing another’s data. And because only compact model updates travel across the network — rather than large-scale data transfers — FL also keeps bandwidth requirements low and aligns naturally with stringent privacy regulations such as the GDPR. One caveat is worth keeping in mind: the model updates themselves must still be protected, since they can conceal traces of sensitive information. This is why practical systems, including PRESERVE, layer additional privacy-enhancing technologies such as secure aggregation on top of plain FL.

Figure 2: Example of a centralised Federated Learning architecture, in which a central server manages model aggregation and updates).

PRESERVE is a real-world lifeline for law enforcement

The PRESERVE project takes these ideas to the next level, applying FL to help law enforcement agencies collaborate across borders. Police case data is among the most sensitive information around, and in many countries, it’s legally barred from leaving home premises. FL solves this by letting each agency train the model locally and only share privacy-protected updates. No raw evidence, no legal headaches, just smarter AI for fighting crime, safely.

PRESERVE combines different mechanisms to achieve reinforced privacy preservation as follows:

Extra safeguards like differential privacy ensure updates can’t be reverse engineered to leak sensitive details.
Each agency retains full control over its data while still contributing to a global cause.

The result is a single, smarter AI model that has effectively learned from cases across many countries, without any raw evidence ever being transferred.

Challenges on the road: Data diversity and security risks

Of course, it’s not all smooth sailing. Practical FL must deal with differences in device power, data formats, network speed, and, critically, the unique patterns in each agency’s data. These differences can lead to weight divergence (where updates clash) or catastrophic forgetting (where local models lose shared knowledge). Solutions like personalised FL and robust aggregation are helping to keep the learning both reliable and fair.

Security is also front and centre. The danger is twofold: hackers stealing data and malicious actors tampering with the model itself. Techniques like zero-knowledge proofs let participants prove their updates are legitimate, without spilling any privacy.

FL is in your pocket!

If this all sounds futuristic, here’s a fun fact: federated learning is already in your life. Google’s Gboard keyboard and Apple’s Siri use FL to learn from your habits without ever sending your private data to their servers. Hospitals are already using FL to train cancer detection models, showing just how practical and powerful this approach has become.

Looking Ahead: The new standard for private, collaborative AI

FL is not a lab experiment waiting for its moment. It is a proven, everyday technique, quietly running behind autocorrect and smart replies on billions of devices right now. The next step is bringing that same idea to domains where the data is even more sensitive, and the stakes for getting privacy right are even higher. Everything we’ve covered so far, the privacy problem, the federated fix, the mechanics, the benefits, and the PRESERVE project points to the same conclusion: FL is no longer just a research idea. It is quietly becoming a working standard for how organisations that cannot share raw data can still collaborate to learn from each other.

Federated learning is an essential learning paradigm that is becoming an everyday technology. Projects like PRESERVE show, it’s paving the way for organisations with sensitive data to collaborate and innovate, all while keeping privacy intact.