Federated Learning

QuanCrypt-FL introduces a secure and efficient federated learning framework that combines fully homomorphic encryption, low-bit quantization, dynamic pruning, and mean-based clipping to defend against gradient inversion and membership inference attacks. The proposed approach significantly reduces communication and computational overhead while maintaining model accuracy. Extensive evaluations on MNIST, HAM10000, CIFAR-10, and CIFAR-100 demonstrate substantial improvements in efficiency, including up to 9× faster encryption, 16× faster decryption, 3× faster training, and 1.5× faster inference compared with existing secure federated learning methods.

BART-FL is a lightweight defense framework for secure federated learning that detects and filters malicious client updates before aggregation. By integrating PCA, cosine similarity, K-means clustering, and a multi-metric statistical voting mechanism, BART-FL effectively defends against backdoor attacks while preserving model accuracy and computational efficiency across multiple benchmark datasets.

A hybrid privacy-preserving algorithm combining Pre-Aggregation Similarity Measurement (PA-SM) and Differential Privacy (DP) enhances the security and resilience of Federated Learning (FL) in Intelligent Transportation Systems (ITS) by defending against adversarial attacks while maintaining model performance.

A hybrid privacy-preserving algorithm combining Pre-Aggregation Similarity Measurement (PA-SM) and Differential Privacy (DP) enhances the security and resilience of Federated Learning (FL) in Intelligent Transportation Systems (ITS) by defending against adversarial attacks while maintaining model performance.

This work presents an empirical evaluation of secure federated learning techniques for autonomous vehicle applications. Using the LISA traffic light dataset, it compares secure aggregation and multiparty computation methods against poisoning and inference attacks, demonstrating how privacy-preserving aggregation can improve the resilience of federated learning systems during both training and inference.