Unreported position data of fishing vessels significantly hinder the accurate identification of their behaviors. AIS (Automatic Identification System) data—a self-reporting tool used for vessel tracking—is often incomplete because IUU fishing vessels deliberately turn it off to evade detection. In this paper, we propose BlindFG, a novel framework for unreported fishing gear classification that leverages a fishing-conditioned latent space to overcome the limitations imposed by AIS-off periods. BlindFG integrates a Fishing-Conditioned Generative Model (FCGen) with a Fishing Context Sommelier (FCSom) to both reconstruct missing trajectory data and classify vessel fishing types under partial AIS availability. FCGen employs a conditional variational architecture, incorporating fishing context embeddings derived from one-hot encoded fishing gear labels to guide the generation of AIS-Off trajectories. FCSom utilizes the completed trajectories within a 1D convolutional classification framework. By aggregating candidate predictions across multiple fishing type hypotheses, the system effectively discriminates among various fishing behaviors, even in the presence of partial AIS data. Experimental results on global datasets demonstrate that BlindFG significantly improves the accuracy of fishing type classification and offers a robust solution for enhancing maritime situational awareness and combating IUU fishing activities. Experimental evaluations demonstrate that our method significantly enhances the prediction of vessel trajectories and the accurate classification of AIS-off fishing practices worldwide.