Apple releases PPML workshop recordings Apfelpatient

Apple has released four selected presentation recordings and a research overview from its in-house workshop on privacy-friendly machine learning. The two-day event brought together internal researchers and external scientists – and reveals where Apple currently sees the open questions surrounding AI privacy.

Apple has published a recap of the 2026 Workshop on Privacy-Preserving Machine Learning & AI on its machine learning blog. The recap focuses on four recorded presentations, a summary article, and references to 24 scientific papers presented at the workshop. The event aligns with a strategy Apple has consistently pursued since the launch of Apple Intelligence: enabling AI capabilities without the privacy compromises that other providers make.

Three thematic focuses of the workshop

The workshop was divided into three thematic tracks: private learning and statistics, foundation models and privacy, and attacks and security. This division reflects the fault lines along which research on privacy-friendly machine learning currently operates – from mathematical theory to the concrete attack surface of models used in production.

Apple itself describes the breadth of topics with keywords such as federated learning, statistical learning, trust models, attack vectors, privacy accounting, and the particular challenges that Foundation Models bring with them. The latter is the most relevant point for the current product strategy: Large AI models, such as those behind features like Apple Intelligence, naturally have a different privacy risk profile than classic machine learning methods.

Four presentations in detail

Apple selected four key presentations from the program and published them in full on its platform. Three of these come from academia, and one from an internal Apple researcher.

The Apple contribution, titled "Crypto for DP and DP for Crypto," is by research scientist Kunal Talwar. It explores the points of contact between cryptography and differential privacy—two disciplines that often need to be combined in practice, but are theoretically distinct worlds.

Three further presentations come from the academic world: Aleksandar Nikolov from the University of Toronto will speak about Online Matrix Factorization and Online Query Release, Elissa Redmiles from Georgetown University about the communication of security and privacy technologies for responsible data collection, and Franziska Boenisch from CISPA about Memorization in Foundation Models – i.e., the question of how and why large AI models sometimes reproduce training data verbatim.

24 scientific papers as the basis for the program

Beyond the lecture recordings, Apple lists 24 scientific papers that were presented at the workshop. Three of these are by current or former Apple researchers and are available directly on Apple's machine learning page.

One of the three Apple papers – "Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem" – is particularly insightful because it directly relates to Apple’s product strategy. Homomorphic encryption allows computations to be performed on encrypted data without decrypting it. A second paper deals with efficient privacy loss accounting for subsampling, and a third with trade-offs in data storage due to so-called strong data processing inequalities. All three topics are closely aligned with the technical challenges Apple must address when scaling Apple Intelligence.

Apple's research signal to the outside world

The publication of the recordings is part of a broader communication strategy. Apple's machine learning blog has been a deliberate counterpoint to the black-box perception of its own product development for years. While Apple is traditionally reserved when it comes to product announcements, it publishes continuously on this research channel.

This makes strategic sense: Apple is in open competition with research teams at Google, Meta, and OpenAI, which have maintained a strong presence in academia for years. Anyone who wants to be taken seriously in AI research needs to be visible at conferences and cultivate collaborations with universities. That's precisely what happens at workshops like these.

Connection to Apple's intelligence strategy

Privacy-friendly machine learning is not an abstract research topic for Apple. The company explicitly positions Apple Intelligence through its privacy model: local processing where possible, private cloud computing with externally auditable code for more complex queries. Methods such as differential privacy, federated learning, and homomorphic encryption form the technical backbone of this promise.

At the same time, the research agenda reveals that Apple is still working on fundamental questions regarding some components. This aligns with the recent discussion surrounding its own AI infrastructure, whose utilization is said to have fallen short of expectations. Anyone who fulfills privacy promises without compromising the performance of their models is performing a technological balancing act – the work presented here sheds light on precisely this interface.

Privacy research as a visible pillar of Apple's AI

By publishing the lecture recordings and the list of scientific papers, Apple is making part of its research pipeline public. For the wider Apple community, this is primarily a signal: Privacy-friendly AI remains a central pillar of its strategy, and Apple continues to invest in the mathematical and cryptographic foundations that will support this claim. (Image: Shutterstock / amgun)