Adaptive Federated Learning with Functional Encryption: A Comparison of Classical and Quantum-safe Options

Federated Learning (FL) is a collaborative method for training aggregate machine learning models while preserving the confidentiality of individual participant training data. Nevertheless, FL is vulnerable to reconstruction attacks exploiting shared parameters to reveal private training data. Cryptographic techniques applied to mitigate this threat either incur high computational cost, require sharing private keys, or add extra communication rounds among participants.In this paper we apply Multi-Input Functional Encryption (MIFE) to a recent FL implementation for training Deep Learning-based network intrusion detection systems. We assess both classical and post-quantum solutions in terms of memory and computational overhead. We find that post-quantum algorithms are more computationally efficient in selective security settings but require considerable memory in adaptive security settings.