Enhancing Federated Learning with Homomorphic Encryption and Multi-Party Computation for improved privacy

Federated Learning (FL) has demonstrated substantial promise in distributed machine learning by enabling collaborative model training without sharing raw data. However, FL systems still face challenges related to privacy and security. This paper evaluates the impact of incorporating Privacy-Enhancing Techniques (PETs) into a Federated Learning architecture conceived for Network Anomaly Detection.This work provides two-fold contributions: the impact of Homomorphic Encryption and Multi-Party Computation on model classification, considering various types of attacks, validated using a widely known dataset, the CIC-IDS2017 dataset, and the performance and resource analysis when compared to a classical Federated Learning architecture (without using PETs).The results show that using Multi-Party Computation presents a residual increase in computational resources (from 0.998MB to 1.417MB of RAM consumption, on average) accompanied by a significant increase in privacy for the different agents involved and even a slightly increased performance achieved by the detection module. Such results offer insights into the trade-offs between security benefits and computational overhead, providing valuable guidelines for optimizing federated learning systems with integrated privacy protections.