In this paper, we investigate a marine Internet of Things (M-IoT) network with high altitude platform (HAP) secure computation offloading at risk of eavesdropping. To ensure the security of HAP’s information transmission, we utilize a group of unmanned surface vehicles (USVs) with an HAP to form a non-orthogonal multiple access (NOMA) transmission group to provide co-channel interference. Our goal is to minimize the total energy consumption by jointly optimizing HAP’s computation offloading workload, HAP’s transmission power, data transmission time, and USV’s transmission power while meeting the security and delay requirements. Although this problem is strictly non-convex optimization, we use problem transformation and vertical decomposition methods to decompose the problem into an underlying problem and a top-level problem. The underlying problem is to determine the HAP’s transmission power and HAP’s computation offloading workload, and the top-level problem is to determine the data transmission time, which is solved by using proximal policy optimization (PPO). The two subproblems are solved iteratively over each other to obtain the minimum total energy consumption. Simulation results show that the proposed algorithm converges faster and reduces the energy consumption of 30.34% compared to asynchronous advantage actor-critic (A3C).