An Investigation of Model Poisoning Dynamics in Federated Learning Scenarios
R Megha, Abhijith A Thampi, Vanapala Bharath Chandra Varma, Greeshma Sarath · 2024
Federated learning was developed as a distributed learning paradigm to train machine learning models using data collected by individual devices without uploading it to a central server. FL safeguards user privacy because the training is conducted on distinct edge devices. Federated learning is susceptible to several attacks, though. Attackers may introduce several types of attacks while transmitting global model parameters and local changes. This study looks into the effects of several model poisoning attack kinds. We apply federated learning using a heterogeneous dataset, concentrating on model updates and adding controlled noise to mimic adversarial situations. The corruption ratios we implement here for analysis are 0.2,0.5, and 0.8. For every poisoning technique, accuracy is rigorously examined, revealing unique effects on model integrity. The results provide deep insights into the susceptibility of federated learning models to adversarial attacks, contributing to the ongoing discourse on the security of collaborative machine learning systems.