The digital world has brought unprecedented convenience and connectivity but also raised significant concerns about data privacy. As we share more of our lives online, the need for robust privacy-enhancing technologies has become paramount. On-device learning has emerged as a powerful tool to protect personal data while enabling advanced capabilities. In this blog, we will explore on-device learning, its role in enhancing privacy, and how it’s used.

What is On-Device Learning?

On-device learning, sometimes referred to as federated learning, is a machine learning approach that allows training models directly on a user’s device with data available on their device. Only updated model parameters are sent to a remote server or cloud. This means that a user’s smartphone, tablet, or other device can learn and adapt to their preferences without constantly sending their data to remote servers. This gives users more control over their data, protects their privacy, and reduces the need to send raw individual user data to external servers.

How does On-Device Learning Work?

On-device learning operates with the following four principles:

1. Local Data Processing: Instead of sending your data to the cloud, on-device learning processes data directly on your device. This can include training machine learning models, recognizing patterns, or adapting to a user’s  preferences.
2. Privacy-Preserving Algorithms: Privacy-preserving algorithms ensure that only the updated model parameters leave the device. The user’s personal data remains on their device and is never exposed to third parties. 
3. Personalized User Experience: On-device learning allows a user’s device to provide a personalized user experience by understanding their preferences, habits, and requirements without compromising data privacy.
4. Offline Functionality: Due to local data processing, on-device learning enables a user’s device to adapt to their preferences immediately even when it's not connected to the internet. This ensures that the user can benefit from personalized features when they’re offline as well.

How are Marketers Using On-device Learning? 

With on-device learning, online retailers can gain insights on consumers’ preferences and behaviors without tracking their individual preferences. The way this works is, each consumer’s device downloads the current model, improves it by learning from the data on their phone. The model updates from each of these devices are then collected, compiled, and are fed back into and improved on the central model. Thus, the marketers just learn the overall purchase pattern or behavior without ever learning individual consumer preferences or behaviors. 

Let’s look at a real-world example of a data collection sequence that uses on-device learning:

1. A user’s web browser downloads a cross-sell prediction model from an advertising platform like Meta ads or Google Ads. 
2. The user clicks an ad and makes a purchase. Let’s say they clicked an ad for a smartphone and subsequently bought a smartphone as well as a screen protector. 
3. The model performs inferences from the purchase data without sending the data to the advertising platform server or cloud. 
4. The model gathers such inferences across millions of devices and compiles them to improve the advertising platform's central model. 
5. Over time, the model improves and can be used to find an increasingly specific audience for screen protectors.

On-device learning is not perfect from a privacy perspective. When model parameters leave users’ devices they still leak information about the underlying local training data. So, the risk of sensitive information being shared is only reduced and not completely eliminated.To prevent this, on-device learning is often combined with other PETs such as differential privacy and secure computation, which we will cover in different posts on our blog.

‍