The obvious solution to AI privacy risk is to not train on sensitive data. This works until you realize that sensitive data is often exactly the data that would make the model most useful. A medical AI trained only on public information is less capable than one trained on real patient records. A fraud detection model trained on anonymized transactions is weaker than one trained on actual fraud patterns. The data you most need to protect is often the data you most need to use.
This tension has driven the development of a family of techniques that share a common goal: letting AI systems learn from sensitive data without that data ever being directly exposed. They work in very different ways, but understanding them together is more useful than encountering each one in isolation.
Federated Learning keeps the data where it lives. Instead of moving training data to a central server, the model travels to the data. Each participant trains a local copy of the model on their own data, then shares only the model updates — not the underlying data — with a central coordinator. A hospital network can collectively train a model on patient records from dozens of institutions without any institution's records ever leaving its own systems.
Differential Privacy adds carefully calibrated noise to data or model updates, making it mathematically impossible to determine whether any specific individual's information was included in the training set. The model learns the statistical patterns in the data without those patterns being traceable back to specific people. The tradeoff is some reduction in model accuracy, and the art is in calibrating how much noise to add.
Homomorphic Encryption takes a different approach: it allows computation on encrypted data without ever decrypting it. A model can process an encrypted input and return an encrypted output, and only the user with the decryption key can read the result. The provider never sees the data in plaintext. This is computationally expensive, but for high-sensitivity applications the cost is often worth it.
Secure Multi-Party Computation (SMPC) allows multiple parties to jointly compute a result using their combined data without any party revealing their individual inputs to the others. It is the cryptographic equivalent of several organizations pooling their knowledge without sharing their secrets.
Privacy-Preserving Machine Learning (PPML) is the broader field that encompasses all of these techniques and the research into combining them effectively. In practice, real deployments often use several approaches together: federated learning to keep data local, differential privacy to protect individual records within the federated updates, and secure aggregation to protect the update process itself.
The articles in this section cover each technique in detail. The common thread is that privacy and capability are no longer as opposed as they once seemed.
