Does Federated Learning Improve the Resiliency of Machine Learning Models Against Data Fault Attacks?

Machine learning (ML) is increasingly used in various fields, making it vulnerable to cyberattacks like label-flipping attacks, which alter training data labels to mislead models. This paper evaluates the impact of such attacks on popular ML models, including Multi-Layer Perceptron (MLP), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN), using two datasets. Results show that while RF is most resistant, MLP performs best under normal conditions but suffers greatly under attack. To mitigate this, we propose a federated learning (FL) approach that distributes training across multiple servers, enhancing robustness. Our results on the Breast Cancer Wisconsin dataset demonstrate FL’s effectiveness, achieving no accuracy drop with 20 servers, a training time of 99.79 seconds, and 3.82 MB bandwidth usage. FL proves highly resilient, making it suitable for critical applications like medical diagnosis in adversarial environments.