Beyond The Data Wall - Federated Learning Without Sacrificing Privacy

There has been a refrain in the business world for years: gather more data, put it in storage, run the analysis, and then do it all over again with even more. 

In some quarters it was like a sport. Every sensor update, support ticket, machine reading and customer click was funneled into an ever-expanding database. And to be fair, the approach had its merits. For a time, bigger datasets were the surest way to build better AI.

Then reality showed up.

Privacy concerns on the part of customers have mounted. Governments have put in place tighter rules. Security departments are no longer able to turn a blind eye to the risks of a centralized data warehouse. 

The notion of funneling everything into one big repository has lost its luster. Put it this way: you could ask everyone at a neighborhood barbecue to empty their refrigerator so you can plan the menu. You might get a fine meal out of it, but most would decline.

Today’s companies are in the same position. They require precise predictions and more capable AI, yet they see little value in having all the raw data in one spot. 

This is what has made federated learning such an elegant and practical solution for enterprise artificial intelligence. Rather than haul the data to the AI, you send the AI to the data.

What Is Federated Learning?

In simple terms, federated learning is a form of machine learning that trains AI models across a number of devices or entities without the data having to move from where it is. 

There is no need to copy sensitive material into a central storehouse. Each system in the network will train its own local version of the model and return only the parameters (the improvements) to a central point for a stronger global model. 

The financial transactions, medical files and customer records stay put. It makes all the difference when it comes to security of data - the less data moves, the better.

Why Traditional AI Hits a Data Wall

The problem with conventional machine learning is that it is predicated on the idea you can corral data in one place. Modern enterprises don’t work like that.

Data is scattered among:

1. Corporate offices
2. Cloud platforms
3. Hospitals
4. Retail locations
5. Manufacturing facilities
6. Mobile devices
7. Partner organizations

Each area has its own ownership issues and security protocols. To try and bring it all under one roof is to invite risk and complicate compliance with GDPR, HIPAA and other regional statutes, not to mention making the whole operation a tempting mark for an attacker.

Federated learning sidesteps that entirely.

Enterprises Businesses Now See It

Enterprise businesses are taking note because the benefits of federated learning goes beyond privacy. It opens up new avenues for collaboration. 

Banks can harden their fraud detection while leaving customer finances alone. A consortium of hospitals can refine disease detection without exchanging patient files. Retailers can make their recommendation engines smarter across the board without building a monolithic warehouse of purchase histories. 

The question is no longer “Is it legal to share this?” but “Can we share what we have learned?

Some Key Business Advantages 

Federated learning offers several practical benefits:

1. Better protection of sensitive customer data
2. Reduction of regulatory compliance risks
3. Lower exposure to centralized data breaches
4. Better collaboration across departments or organizations
5. Continuous improvement of business models using distributed data sources
6. Improved AI performance in geographically spread out operations

None of these advantages eliminate security concerns, but they dramatically reduce the amount of sensitive information that needs to move between systems. This along results in better security.

Manufacturing: Smarter Predictive Maintenance

Take manufacturing as a case in point. Predictive maintenance is nothing new, and all their manufacturing systems collect a lot of data, and the operational data from sensors tracking pressure, vibration and energy use is valuable. 

Yet factories tend to be independent operations with their own networks and policies; in a supply chain they may even be different companies. With federated learning, each plant can train its maintenance model in house. They contribute to the overall accuracy of the model without uploading years of production data to some central server.

For a firm with hundreds of sites around the globe, the ability to foresee equipment failure and cut down on downtime is more than a marginal gain. It is a competitive edge.

An Interesting Balancing Act In Retail

There is a certain tension in retail today. On one hand, the customer wants a shopping experience that is tailored to them; on the other, they want their privacy to be respected by the business. The two do not always sit well together.

Federated learning has emerged as a sensible way to bridge that gap. Instead of sending full purchase histories to some central AI for processing, local models can pick up on shopping patterns right there in the regional system or on the customer’s own device. What the central model gets are updated parameters for its learning, nothing more than that.

Practical Retail Applications

We see federated learning being put to work in Retail for:

1. Product recommendations
2. Dynamic pricing
3. Inventory forecasting
4. Customer segmentation
5. Demand prediction
6. Personalized promotions

It allows for a level of recommendation quality that does not come at the expense of privacy - something that is at a premium these days as it becomes harder to win over consumer trust.

The Case For Healthcare 

Perhaps no sector illustrates the merit of federated learning like healthcare. Hospitals are sitting on vast troves of medical data, yet when it comes to sharing patient records to better outcomes, privacy law is a major obstacle.

Federated learning alters the dynamic. A hospital can train an AI on its own patient files and then combine those model updates with others to build a superior diagnostic tool for all parties involved. 

The patient file never leaves the building. 

We have seen this prove effective in areas from disease prediction to medical imaging. It is a way for even a small regional hospital to tap into the knowledge of a much larger network.

Financial Services and Fraud Detection

Banks are in a perpetual fight with fraudsters who devise new attack vectors faster than a static model can keep up. 

In the past, the solution was to pool huge quantities of transaction data. But with strict regulations and customers’ justifiable expectations of confidentiality, that is less tenable.

Through federated learning, several institutions can make their fraud detection more robust without ever having to exchange account records. They contribute to the model based on what they see locally. Sensitive information is kept private, but the fraud is all the easier to spot.

Technical Challenges Businesses Should Understand

Federated learning is not a magic bullet, however. There are some real engineering realities to contend with, and companies need to take them into consideration:

Communication Overhead

Communication overhead is one. Coordinating the training of a model across thousands of devices is no small matter; poor bandwidth or an intermittent connection will put a brake on model updates. 

Data Differences

The matter of data diversity should not be overlooked, but most of the time is it - to the detriment of all involved. 

The operational data coming out of a plant in Texas is not going to look like what you find coming out of a plant in Germany, and hospitals have their own specializations for each location as well. 

That makes for uneven local datasets and a more difficult job of producing an accurate global model.

Security Risks

And while the raw data is secure, there is always the risk of an attacker trying to tamper with model updates. Most businesses will layer on extra protections like differential privacy and secure aggregation to mitigate that.

Building a Federated Learning Strategy

A company looking to adopt federated learning should not view it as just another AI initiative. It calls for input from several areas:

1. Legal
2. IT
3. Cybersecurity
4. Data Scientists 

A sound strategy involves pinpointing which data must stay put and setting the proper governance and security controls.

The most successful deployments start with a tightly scoped pilot to work out the kinks in the process before taking it enterprise-wide.

The Future of Enterprise AI Is Becoming More Distributed

Enterprise AI is on a path to becoming more distributed. With privacy rules in a state of flux, one does not expect to see a return to the days of unbridled, centralized data hoarding by companies. If anything, the trend is in the other direction.

We are seeing artificial intelligence disperse as edge computing and connected devices proliferate, all while business ecosystems grow ever more collaborative. In such an environment federated learning has a natural place. It is a matter of letting organizations put their intelligence to work without having to cede control over their most valuable assets. A fine point, perhaps, but one that could well come to define enterprise AI for the coming ten years.

Conclusion

In the end, there is no longer a need for businesses to pit the development of robust AI against the imperative of safeguarding sensitive information. Federated learning provides a way to have it both ways: to innovate and keep private at the same time. 

Whether for fraud detection or predictive maintenance, or in the realms of retail and healthcare, the technology has made its case across various sectors. There are still technical hurdles to clear in terms of security and data consistency, but improved governance is taking care of those.

For any organization with AI plans on the horizon, the course of action is plain. Look at where your current machine learning efforts are being held back by the nature of the data and see if federated learning can deliver the same results without the need to move it around. 

Those who put privacy first in their architecture now will be in a stronger position down the road to earn trust, stay in compliance and make the most of what comes next.

Please contact us at ScreamingBox if you want to discuss Federated machine learning, and how to integrate it into your development project.

Check out our podcast on Cybersecurity discussing the related topic of Data Privacy and AI.