SuperFL: Bridging LDP With Byzantine Robustness in Federated Learning on Non-IID Data for Low-Altitude Networks

Driven by the low-altitude economy, heterogeneous aerial vehicles (e.g., AAVs, piloted aircraft, eVTOLs) are expected to collaboratively support Artificial General Intelligence (AGI) models. Their data are Non-Independent and Identically Distributed (Non-IID), privacy-sensitive, and vulnerable to Byzantine attacks. Existingriven by the low-altitude economy, heterogeneous aerial vehicles (e.g., AAVs, piloted aircraft, eVTOLs) are expected to collaboratively support Artificial General Intelligence (AGI) models. Their data are Non-Independent and Identically Distributed (Non-IID), privacy-sensitive, and vulnerable to Byzantine attacks. ExistingD Federated Learning (FL) schemes still struggle to efficiently mitigate both privacy leakage and Byzantine threats in low-altitude networks characterized by mobility and data heterogeneity. We propose SuperFL, a secure FL framework that bridges Local Differential Privacy (LDP) and Byzantine resilience for Non-IID low-altitude networks. In each round, every device (i.e., client) splits the model into small segments and compresses them to perform parallel aggregations. The compressed sensing drastically reduces the number of parameters subjected to LDP perturbation, cutting the volume of noise while preserving geometry. Additionally, each client adaptively tunes LDP noise on the compressed segments to optimize aggregation accuracy while ensuring privacy, and shuffles those noisy segments to curb privacy-budget inflation from repeated rounds. Owing to approximate invariance of cosine similarity and norms under compression, the server can effectively identify and filter Byzantine clients in Non-IID data settings via executing the cosine-similarity-based robust aggregation rules on locally perturbed and aggregated model segments under compression without significant accuracy losses. The privacy and correctness of SuperFL are theoretically analyzed. Experimental results confirm its effectiveness and efficiency in low-altitude networks.