Abstract: Unmanned aerial vehicle （UAV） communication is a key enabling technology for information exchange among low-altitude network nodes， and ensuring its security is critical for the healthy and stable development of the low-altitude economy. Physical layer key generation （PLKG） leverages the randomness and reciprocity of wireless channels to dynamically generate shared keys， thereby enhancing system security and privacy protection. This approach has particular advantages in resource-constrained application scenarios， such as small UAV platforms， where storage and computational resources are limited. However， the high-speed movement of UAVs or environmental scatterers introduces Doppler shifts， which cause the channel to exhibit high time variability and consequently affect the consistency of channel random parameters obtained through bidirectional channel estimation. Recent studies have shown that， compared with conventional modulation schemes such as orthogonal frequency division multiplexing （OFDM）， orthogonal time-frequency space （OTFS） modulation can better handle fast time-varying channels， making it particularly suitable for high-mobility communication scenarios and applicable to physical layer key generation. To address the challenge of physical layer key distribution in high-mobility scenarios， this study proposes an OTFS-based key generation scheme. The proposed scheme can effectively cope with fast time-varying channels and reduce the bit disagreement rate （BDR） in key generation. For accurate channel parameter acquisition in the delay-Doppler （DD） domain， which is crucial for subsequent steps， a sparse Bayesian learning （SBL）-based channel estimation method is adopted. Traditional estimation methods struggle with computational complexity and the off-grid issue caused by non-integer delays and Doppler shifts. The SBL approach， formulated within a variational inference framework， can effectively estimate the sparse DD-domain channel impulse response， including path gains， delays， and Doppler shifts， with improved accuracy， especially in noisy conditions. Since the slow variation of OTFS channels may cause insufficient randomness and low entropy in the generated keys， the discrete Karhunen-Loève transform （KLT） is employed to decorrelate channel components. Furthermore， to enhance the key generation rate， a joint quantization algorithm is proposed in the path coefficient-Doppler-delay domain. A multi-bit adaptive quantization （MAQ） scheme， guided by the transmitting node， is implemented to ensure high bit agreement between legitimate users while minimizing information leakage. Simulation results show that the scheme achieves strong performance in improving key generation rates， optimizing entropy characteristics， and reducing bit disagreement rates. Specifically， the proposed scheme can effectively adjust performance metrics， such as the bit disagreement rate and key generation rate， by setting a bit disagreement rate threshold， thereby achieving a dynamic balance between security and efficiency while maintaining higher key entropy. This strong randomness significantly increases the time and computational resources required for attackers to perform brute-force attacks or other forms of cryptographic analysis. The proposed framework provides an efficient and lightweight solution for secure key generation in high-mobility UAV communications. It effectively addresses challenges posed by channel time variability， slow variation， and sparsity in the DD domain. By integrating OTFS modulation， SBL-based channel estimation， KLT-based decorrelation， and joint multi-parameter quantization， the scheme ensures a high key generation rate， low bit disagreement， and high entropy， making it a robust candidate for enhancing the physical layer security of next-generation low-altitude networks. The core ideas also hold potential for extension to future 6th generation （6G） scenarios involving millimeter-wave and massive multiple-input multiple-output （MIMO） systems.