False Data Injection Attacks Against UAV Battery State of Health Estimation with A Capacity-informed Latent Diffusion Model

In low-altitude unmanned aerial vehicles, reliance on battery state of health (SOH) estimation on wirelessly sensed data introduces a severe vulnerability to false data injection attacks, which leverage manipulated data to mislead computing and control systems. To underscore the security vulnerabilities of batteries in the emerging low-altitude economy, this work pioneers the application of an advanced latent diffusion model to synthesize battery data and induce the SOH estimation model to yield deceptive results. To achieve this purpose, we propose a conditional variational autoencoder-based latent diffusion model (vLDM). First, the diffusion-based denoising process operates within a compact latent space learned by a variational autoencoder, ensuring high-fidelity generation of voltage profiles. Second, to maximize attack effectiveness, we further design a prediction-guided training paradigm that guides the vLDM using the target model’s SOH predictions for a known battery to learn the target model’s specific capacity-to-voltage inverse dynamics. Third, the proposed model achieves rapid attack execution by generating high-fidelity voltage profiles in just a few latent sampling steps, facilitated by robust initial reconstruction from the variational autoencoder architecture. Finally, experimental validation on a public electric vertical takeoff and landing dataset demonstrates the proposed vLDM’s high effectiveness and rapidity, achieving an average R 2 score over 0.97 in misleading the estimator with adversarial samples generated in as few as four latent steps. Popular detectors fail to identify the attacked voltage cycling data when tested under the proposed attack paradigm.