Data Poisoning Attacks on Federated Machine Learning

Federated Learning (FL) is a privacy-preserving machine learning paradigm that enables decentralizeddevices to collaboratively train a global model without sharing raw data. However, this decentralizedsetup introduces new vulnerabilities—especially to data poisoning attacks, where adversarial clients injectmaliciously crafted data into their local training to degrade the global model’s performance or causetargeted misclassifications. This paper presents an in-depth analysis of data poisoning attacks in FLsystems and proposes a robust detection and mitigation framework to address them. Various poisoningstrategies, such as label flipping and backdoor attacks, are studied, and their impacts on federated learningare quantified. We evaluate the performance of the proposed defense using benchmarks like MNIST andCIFAR-10 under different attack intensities. Experimental results show that our framework caneffectively identify poisoned updates with over 90% accuracy and restore model performance close to theclean baseline, highlighting the critical need for resilient aggregation strategies in federated learningdeployments.