Neuro-Symbolic Federated Learning with Secure Multi-Party Computation

In some critical applications, multiple parties want to train a machine learning algorithm (e.g., a neural network) on sensitive data. This sensitive must remain secret to the respective parties.

Recent research on federated learning has rendered distributed learning with confidential information practical. This project researches the feasibility of neuro-symbolic learning—i.e., techniques that combine learning from data with logical constraints—using secure multi-party computation—i.e., techniques that enable the computation of a public function while keeping the inputs private.

Neuro-symbolic learning is known to help address some issues with the use of machine learning algorithms in critical applications, including the ability to maintain logical consistency, and to achieve robustness against adversarial attacks. With secure multi-party computation, the benefits of neuro-symbolic learning would be applicable to new domains where training data cannot be shared.