Editorial illustration for Apple Workshop Shows ML with Homomorphic Encryption, Georgia Institute, CISPA
Apple Workshop Shows ML with Homomorphic Encryption,...
Apple ran a workshop last week about doing machine learning without ever seeing the raw data. The trick is a notoriously slow, cryptographically heavy technique called homomorphic encryption, which lets models run calculations directly on scrambled numbers. The academic pitch is that it's a perfect privacy shield.
The practical reality is that it's been too sluggish to use. Researchers from Georgia Tech, CISPA, and Apple itself now say they've wired it up to work inside Apple's own software stack, turning a lab curiosity into something that might actually run.
They weren't the only ones closing gaps. Monika Henzinger and a team from Austria and Harvard figured out how to keep privacy promises from breaking when multiple data analyses happen at once. Google's Lillian Tsai and Eugene Bagdasarian proposed tailoring security rules to what an AI agent is actually trying to do, instead of using one blunt policy for everything. Another group found that strategically removing chunks of training data can make a model better at remembering facts, a weird result that flips standard assumptions about memorization.
Appleâs fundamental research has consistently pushed the state-of-the-art in this domain, and earlier this year, we hosted the Workshop on Privacy-Preserving Machine Learning & AI.
This wasn't a theoretical conference. The work shown here was about building things that function, specifically inside Apple's walled garden. The underlying message from the company was blunt.
Privacy can't be a feature you bolt on later. It has to be part of the machine from the start. They are making it a core engineering problem, not a compliance checkmark.
The trade-off between useful models and private data is being dismantled piece by piece, with math and code.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives OpenAI opens GPT-5.5-Cyber to vetted security researchers, adds three tiers
From Archives LightSeek launches TokenSpeed, cutting LLM latency by half vs TensorRT-LLM
Common Questions Answered
What is homomorphic encryption and how does it enable machine learning on encrypted data?
Homomorphic encryption is a cryptographic technique that allows machine learning models to run calculations directly on scrambled numbers without ever decrypting the raw data. This means the data remains encrypted throughout the entire computation process, providing a theoretically perfect privacy shield for sensitive information used in model training and inference.
Why has homomorphic encryption traditionally been impractical for real-world machine learning applications?
Homomorphic encryption has historically been too slow and computationally heavy to use in practical scenarios, making it unsuitable for production machine learning systems. The cryptographic overhead required to perform calculations on encrypted data significantly slowed down model performance, limiting its adoption despite its strong privacy guarantees.
What did the Apple workshop with Georgia Tech and CISPA researchers demonstrate about implementing homomorphic encryption?
The workshop showed that researchers from Apple, Georgia Tech, and CISPA have successfully integrated homomorphic encryption into Apple's own software systems to make it function in real-world applications. This represents a significant breakthrough in making the previously impractical technique viable for actual machine learning workloads within Apple's ecosystem.
How does Apple's approach to privacy in machine learning differ from traditional security practices?
Apple emphasizes that privacy must be built into machine learning systems from the ground up rather than added as an afterthought or compliance requirement. The company treats privacy as a core engineering problem, using mathematical and code-based solutions like homomorphic encryption to dismantle the traditional trade-off between useful models and private data protection.
Further Reading
- Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem — Apple Machine Learning Research
- Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem — School Information System
- Apple Workshop on Reasoning and Planning 2025 — Apple Machine Learning Research
- STOC 2025 Program — ACM STOC 2025
