Machine learning has become a central tool for extracting insights from data in fields such as healthcare, security, and public services. However, these advances often come with risks to privacy, fairness, and trust. Privacy-Preserving Machine Learning (PPML) seeks to address this challenge by enabling models to learn from data without exposing sensitive information.
What is PPML?
PPML refers to a collection of techniques that make it possible to train and apply machine learning models directly on encrypted or otherwise protected data. Approaches such as homomorphic encryption, functional encryption, differential privacy, and secure multiparty computation are used to ensure that individuals’ data remains private, even during analysis. In practice, this means that a model can still deliver accurate predictions and classifications while protecting the confidentiality of the underlying information.
Why it matters
The use of PPML is particularly relevant in domains where sensitive personal data is involved, such as health records or security-related information. Traditional methods of anonymisation or aggregation are no longer sufficient, as advances in re-identification techniques can still compromise privacy. PPML offers a stronger foundation by combining cryptographic methods with privacy-by-design principles.
HARPOCRATES and PPML
In HARPOCRATES, PPML is a key enabler of the project’s goal to support federated data sharing and analysis for social utility. The consortium is developing privacy-preserving models that can:
-
Classify encrypted data with high accuracy
-
Operate across federated data spaces, where data remains within the control of its owners
-
Eliminate reliance on biased proxies such as geography, gender, or race
By embedding PPML into its architecture, HARPOCRATES seeks to demonstrate how large-scale analysis can be performed without undermining citizens’ rights to privacy and data protection.
Real-world use cases
HARPOCRATES applies PPML to two cross-border pilot scenarios:
-
Health data for sleep medicine – supporting better diagnosis and treatment pathways while keeping patient records protected.
-
Threat intelligence for local authorities – enabling secure information sharing and analysis without exposing sensitive or classified data.
These examples show how PPML can balance the benefits of advanced analytics with the need for strong privacy safeguards.
Looking ahead
PPML is not only a technical innovation but also a cornerstone for building trustworthy digital ecosystems. By advancing and integrating these methods, HARPOCRATES contributes to a future where data can be used responsibly to improve services and decision-making, while maintaining fairness and protecting individual rights.