<p>With the rapid development of trajectory data applications, the risk of privacy breaches has become increasingly prominent. Although existing research has improved data usability through adaptive perturbation or deep generative models, it still struggles to utilize prior distribution information fully under privacy constraints, leading to distribution bias and a decline in data quality. To address this issue, this paper proposes a trajectory generation framework (LGAN-Geo) that combines differential privacy and deep generative models. First, this paper designs a two-stage data collection mechanism. This mechanism estimates the prior distribution of user locations using sample data and constructs a prior-based perturbation method under the constraint of Geo-Indistinguishability, thereby reducing the distribution bias caused by random perturbations and enhancing data usability under the same privacy budget. Secondly, this paper proposes a trajectory encoding method based on Bayesian posterior probability. This method represents the perturbed location as a probability distribution vector of the true location, thereby providing a more expressive input representation for the generative model while preserving the spatiotemporal distribution characteristics of the trajectory. Finally, this paper constructs a trajectory generation model that integrates Long Short-Term Memory networks and Generative Adversarial Networks. This model is used to learn the temporal dependencies and distribution characteristics in privacy-protected data, thereby generating trajectory data with good temporal consistency and distribution similarity. Experimental results show that LGAN-Geo demonstrates competitive performance in terms of data availability, distribution similarity, and resistance to privacy attacks, achieving a better utility-privacy trade-off under privacy protection constraints.</p>