Apple’s Latest Entry in its Machine Learning Journal Focuses on Privacy and Search

Ever since Apple published its dedicated site for its machine learning journal, the company, and the teams within, have been publishing interesting insights into how certain elements work.

We have seen entries on a variety of topics, with two meant to address Siri: One in development of the digital personal assistant, and details regarding the “Hey, Siri” feature. And now the Differential Privacy Team has come forward and penned their own entry, this one focusing on how Apple can try to better the user experience on iOS and Mac devices, while also doubling down on their efforts for customer security.

“Understanding how people use their devices often helps in improving the user experience. However, accessing the data that provides such insights — for example, what users type on their keyboards and the websites they visit — can compromise user privacy. We develop a system architecture that enables learning at scale by leveraging local differential privacy, combined with existing privacy best practices.”

As is usually par for the course, the new entry in Apple’s machine learning journal is on the technical side — it is designed primarily for researchers, engineers, and developers — but it’s certainly worth a look through if you’re even remotely interested in how Apple develops some of its security initiatives.

The team discusses topics like how the iOS and macOS devices can privatize data, making it secure on the device in question without having to worry about sending out data to developers or Apple itself. Apple customers have to opt-in to share analytics with Apple or third-party app developers:

“Users have the option, in System Preferences on macOS or in Settings on iOS, to share privatized records for analytics. For users who do not opt in, the system remains inactive. For users who do opt in, we define a per-event privacy parameter. Additionally, we set a limit on the number of privatized records that can be transmitted daily for each use case. Our choice of […] is based on the privacy characteristics of the underlying dataset for each use case. These values are consistent with the parameters proposed in the differential privacy research community, such as [5] and [6]. Moreover, the algorithms we present below provide users further deniability due to hash collisions. We provide additional privacy by removing user identifiers and IP addresses at the server where the records are separated by use case so that there is no association between multiple records.”

Oh, and if you were curious, the latest journal entry also nots that the tear-face emoji, the one that looks like it’s crying because it’s laughing so hard, is the most used emoji right now. It has a use case of over 25%, with the heart emoji coming in second just over 5%.

The full journal entry can be visited through the source link below.

[via Apple]