Privacy-Preserving Federated Learning with Multi-Key Homomorphic Encryption

Background

Advancements in cryptography, federated learning and platform security enable radical improvements in the security and privacy of data processing. Techniques such as Federated Learning allow to process data in a distributed setting, without aggregating it in a central location. Secure aggregation approaches - such as symmetric homomorphic encryption - allow to aggregate updates from local federated learning nodes without
revealing individual contributions. One such implementation is CanaryBit’s Heflp implementation (https://github.com/canarybit/canarybit-heflp).

However, currently this faces practical limitations because it requires using a shared key for homomorphic encryption among all local nodes.

Objectives

We use the latest hardware security features and privacy enhancing technologies to enable confidential data analytics and collaboration. Our solutions provide strong security and privacy guarantees towards customer
data and workloads throughout the entire lifecycle. To enable privacy-preserving sharing of threat intelligence information, we have developed the HEFLP library as an extension of the Flower Federated Learning
framework, with 3 secure aggregation schemes (CKKS, BFV, and Flashe). An outstanding challenge is to enable secure aggregation of parameters in Federated Learning with multi-key homomorphic encryption. The thesis
includes the following objectives:

- Evaluate the state of the art in the area of multi-key homomorphic encryption schemes;

- Implement secure aggregation for Federated learning using one or more multi-key homomorphic encryption schemes, by extending the existing open-source Heflp implementation.

- Evaluate performance prototype implementation and a written report on the findings.

A successful project could lead to a valuable open-source contribution or an academic publication presented at a prestigious conference or workshop.

Terms

Supervisor: Nicolae Paladi, PhD
Scope: 30 points
Start: Fall Semester 2025, Spring Semester 2026
Compensation: 12 000 SEK upon successful completion of the thesis.

Candidate profile:

We expect you to have good programming skills in: Python/C/Rust as well as familiarity with ML frameworks such as TensorFlow, Pytorch and others. You have an interest in machine learning, cloud computing, security and cryptography and systems performance. Strong oral and written English skills are expected.

Send in your application as soon as possible, by August 27th, 2025 at the latest. Applications will be reviewed on a rolling basis and the position will be closed as soon as we find a good candidate. Note that only complete
applications will be reviewed. To be considered complete, applications must include:

- Your CV with your education, professional experience and specific skills;

- A written report you authored or co-authored for a university-level course;

- Samples of previous programming or other relevant projects;

- Recent grades (academic transcript).

See a list of earlier publication and master’s thesis projects done at CanaryBit: https://www.canarybit.eu/research-and-technological-leadership/