The information technology industry is ever-evolving. We have come a long way from mainframe applications and relational databases to SaaS and mobile applications. Each of these developments offered significant competitive advantages to organizations. Now, artificial intelligence is being hailed as the next important innovation of the industry. Over the last few years, organizations have been rushing to adopt AI solutions to maintain their competitive edge in the market and keep their ability to offer unrivaled services intact.
AI can offer intelligent insights, help organizations make more informed decisions, greatly enhance operational efficiency, and much more. In fact, a recent survey conducted by ManageEngine found that 86% of organizations have increased their AI usage over the last two years. Businesses all across the world have high hopes from AI solutions. However, as with any technology, there are challenges to its implementation and performance.
Regulatory concerns hinder AI progress
One of the main issues with enterprise AI initiatives is the need to aggregate vast amounts of “clean” data from diverse reliable sources and pool this data in one central server, but these requirements raise multiple challenges. To start, data privacy regulatory bodies across the globe have voiced concerns regarding data sharing, usage, and storage practices. Almost every nation has already implemented or is in the process of implementing privacy regulations that have far-reaching effects on the data management practices of organizations.
Existing privacy regulations such as the GDPR and the CCPA are also constantly evolving to accommodate the demands of an increasingly privacy-conscious world. Take the EU-US privacy shield, for example, the privacy shield enabled organizations from the United States of America and the European Union to transfer personal data between themselves. However, this EU-US privacy shield was invalidated by the European Court of Justice in July 2020 on the grounds of inadequate privacy protection.
Most data privacy regulations also have data residency and data localization requirements. This means that personal data will need to remain in the same jurisdiction where it was created and cannot be transferred anywhere else. This presents significant challenges for the development of AI models, the performance of which depends upon this very data.
Privacy by design is becoming a mandate
With the rise of these privacy regulations, embedding privacy into every operation and every workflow of your organization is not only ethical but will also help your business gain a competitive edge over others. There’s a new breed of solutions called privacy-enhancing technologies (PETs) that are designed to help organizations minimize their privacy risk and ensure that their data management practices aren’t flouting any laws.
PETs allow businesses to leverage the ever-increasing amounts of data while guaranteeing personal information stays private. For AI projects, organizations can securely train their models on sensitive data with the help of PETs.
One such PET that takes traditional machine learning a step further is federated learning.
Standard machine learning models require data to be centralized in one server or data center. The entire process of training the model and analyzing the data usually takes place in this central server. For example, if an e-commerce organization wants to predict its customers’ propensity to buy a specific product, it will train its model on data collected from the website and the app.
It can collect thousands of data points about its customers such as products purchased most often, time spent on product pages, products added to wish lists, and more. This data set is then fed back to the central server where it is processed and used for building predictive models. When the model is ready, it is sent back to the devices in order to show the predictions it was made for. The downside of this round-trip process is that it can limit the model’s ability to learn in real-time. It also often requires sensitive data to be sent to a different location (that of the central server).
These challenges are addressed by federated learning. Federated learning takes the machine learning model and brings it to the edge of the network.
Understanding federated learning
In federated learning, the model is downloaded onto the edge device, such as mobile phones, which then develops an updated model using data gathered by the local device itself. The edge device essentially improves the model by learning from the data it has collected. This takes place across multiple edge devices, each of which computes its own version of the updated model. These locally trained models are then sent from the devices back to the central server where they are aggregated. Finally, a single consolidated and improved global model is sent back to the devices.
This means instead of being limited to large, central machines, the model is distributed across mobile devices for real-time computation. Federated learning enables machine learning algorithms to learn from a wide range of data sets sourced from different locations. This approach also makes it possible for multiple organizations to collaborate without needing to share sensitive data with each other directly. The raw training data always remains on the user’s device, and no individual updates are identifiably stored in the cloud.
As several training iterations take place, the shared models can train on a significantly broader range of data than what a single organization can possess in-house. In other words, federated learning decentralizes machine learning by removing the need to aggregate data in a single location.
This decentralization enables faster deployment and testing, lowers latency, and lessens power consumption while ensuring privacy. In addition to this, the enhanced local model on the edge device can be used immediately, powering experiences personalized by the data the mobile device collects.
Real-time use cases of federated learning
Federated learning truly takes AI to its next generation. The enterprise AI landscape today is limited by its inability to train ML models on user-generated data. Organizations have to rely on synthetic data, which is seldom an accurate representation of real-world data.
With federated learning, models can now be built upon huge amounts of real-time data without having to worry about the privacy aspect since the raw data never leaves the edge device. This, in turn, also addresses ethical AI concerns as organizations can go beyond their data boundaries and collaborate with diverse data sources.
This greatly benefits industries, such as healthcare, which are bound by strict privacy constraints, given the nature of data they operate with. With federated learning, healthcare organizations can collaborate to build intelligent models trained upon data collected from multiple healthcare organizations.
One of the biggest potential applications of federated learning can be found in consumer health applications. With the rise in the Internet of Medical Things, there’s been a boom in the usage of healthcare wearables and applications. With many big brands making cardiac health trackers, sleep quality trackers, fall detection systems, and emergency SOS solutions major products of their portfolio, federated learning can enhance the functioning of this segment without risking the safety of personal health information.
Another example of an industry that deals with highly sensitive information is the financial sector. Take the example of a credit scoring agency. Various financial institutions such as banks, card issuers, and auto finance companies send people’s data to credit scoring agencies to request their credit scores. These credit scores are then used to assess a consumer’s default risk, determine if credit should be extended to the consumer, and so on.
This process also requires sensitive financial information to be shared among organizations. However, instead of financial institutions sending data to credit score agencies, the agencies can employ federated learning to generate the credit scores of consumers. The agencies can thus create a single, holistic credit scoring model without having to access consumer data.
Given the age of data explosion that we live in, the applications of federated learning can be many and varied. For the enterprise AI roadmap, this new technology presents a beautiful opportunity to democratize AI. It also opens up new avenues for new applications and provides a novel way of solving large-scale ML problems. In a world where hyper-personalization and highly contextual recommendations are going to help set companies apart, federated learning will play a vital role in the future.
