Privacy-Preserving Training

Banks, hospitals, and telecom providers all face network attacks — but they can’t pool their data to train better detectors. Privacy regulations, competitive concerns, and security policies prevent it. Yet a model trained on one organization’s data may not generalize to another’s network. How do you train a robust classifier across organizations without anyone revealing their data?

What You’ll Work On

This theme explores the intersection of secure multiparty computation (MPC), federated learning, and differentiable logics. The goal is to enable collaborative training where logical constraints improve robustness, and cryptographic protocols protect privacy.

Possible thesis directions:

Feasibility study: Implement a basic federated NIDS training pipeline with MPC (using frameworks like MP-SPDZ) and measure the computational overhead
DL under MPC: Investigate which differentiable logics are tractable under secure computation — some loss functions are much more expensive to compute securely than others
Protocol design: Design new training protocols that balance privacy guarantees, model accuracy, and computational cost
Multi-valued logics for MPC: Explore whether new logics can be designed specifically for efficiency under secure computation

What You’ll Learn

Secure multiparty computation (garbled circuits, secret sharing)
Federated learning architectures
Differentiable logics and neuro-symbolic AI
Privacy-preserving machine learning

Relevant Literature

Supervisors: Alessandro Bruni (ITU), Nicola Dragoni (DTU) – see Team