BART-FL: A Backdoor Attack-Resilient Federated Aggregation Technique for Cross-Silo Applications

Federated Learning (FL) is a decentralized learning method that enables collaborative model training while preserving data privacy. This makes FL a promising solution in various applications, particularly in cross-silo settings such as healthcare, finance, and transportation. However, FL remains highly vulnerable to adversarial threats, especially backdoor attacks, where malicious clients inject poisoned data to manipulate global model behavior. Existing outlier detection techniques often struggle to effectively isolate such adversarial updates, compromising model integrity. To address this challenge, we propose Backdoor Attack Resilient Technique for Federated Learning (BART-FL), a novel lightweight defense mechanism that enhances FL security through malicious client filtering. Our method integrates Principal Component Analysis (PCA) for dimensionality reduction with cosine similarity for measuring pairwise distances between model updates and K-means clustering for detecting potentially malicious clients. To reliably identify the benign cluster, we introduce a multi-metric statistical voting mechanism based on point-level mean, median absolute deviation (MAD), and cluster-level mean. This approach strengthens model resilience against adversarial manipulations by identifying and filtering malicious updates before aggregation, thereby preserving the integrity of the global model. Experimental evaluations conducted on the LISA traffic light dataset, CIFAR-10, and CIFAR-100 demonstrate the effectiveness of BART-FL in maintaining model performance across diverse FL settings. Additionally, we perform a comparative analysis against existing backdoor defense techniques, highlighting BART-FL’s ability to improve security while ensuring computational efficiency. Our results showcase the potential of BART-FL as a scalable and adversary-resilient defense mechanism for secure training in cross-silo FL applications.