In collaborative learning (CL) – for instance, federated learning and split learning – multiple clients collaboratively train a model while ensuring their data remains local and private. However, traditional CL has two severe issues. The first issue is that the privacy promises of traditional CL have been broken where an adversary can launch various types of attacks to either reverse-engineer the client data or infer sensitive properties of such data even when all client data is kept local. The second issue with traditional CL is that its communication overhead is significant, especially with a model of ever-growing size and with a large number of participating clients. Therefore, the task of preserving privacy while ensuring efficient communication and computation is a fundamental challenge in CL.

This internship is to explore collaborative learning, differential privacy, as well as potential systems mechanisms like gradient compression. You will work towards creating a CL system that works for even large ML models, with tunable differential privacy and efficiency guarantees that self-adapt to the user needs and the underlying infrastructure constraints. Ideally, this project will lead to a publication at a top academic venue.

• Student enrolled in a PhD program in Computer Science/Engineering.

• Strong programming skills in Python and ML systems.

• Experience in designing, implementing and evaluating distributed systems is a big plus.

• A strong publication record is a big plus.

• Language skill: English.

Location: Stuttgart (Germany)

• You will be expected to get up to speed with various collaborative learning schemes, as well as differential privacy and ML system efficiency mechanisms.

• You will design a novel collaborative learning system that achieves both differential privacy and resource efficiency, even for large ML models.

• You will implement a working prototype and be involved in writing an academic paper related to the project.