In fog computing environments, task offloading mechanisms are widely used because they can effectively alleviate the computing pressure of terminal devices. However, with the increasing demand for data privacy protection and the rise in network attack risks, achieving efficient task scheduling while ensuring data security has become a key challenge. Current research focuses on improving offloading efficiency, often ignoring the security threats that may be faced during task allocation and execution, and has certain limitations. Therefore, we propose a deep reinforcement learning task partial offloading (DRLPO) mechanism suitable for multi-user and multi-server scenarios, which divides the application device task into k subtasks, and models the offloading problem of each subtask after block as a Markov decision process (MDP) for distributed processing. By offloading subtasks instead of complete tasks, information leakage is prevented. In addition, we model the task offloading problem as a multi-objective optimization problem that takes into account energy consumption and delay, and use the Attention-DQN mechanism to keep the weighted sum of task processing delay and energy consumption at the lowest level. Experimental data show that compared with the existing baseline algorithm, the average task processing time and energy consumption of DRLPO are reduced by 2.2 s and 1.1 KJ, respectively, which can effectively reduce the response time of device tasks.