Journal of Cybersecurity and Information Management JCIM 2690-6775 2769-7851 10.54216/JCIM https://www.americaspg.com/journals/show/3771 2019 2019 Research scholar, Department of Computer Science and Engineering, Faculty of Engineering, Karpagam Academy of Higher Education, Coimbatore, Tamil Nadu, India R. R. Professor and Head, Department of Computer Science and Engineering, Faculty of Engineering, Karpagam Academy of Higher Education, Coimbatore, Tamil Nadu, India R. Santhosh Multi-cloud computing is emerging as a transformative solution to meet the extensive computational demands of Internet of Things (IoT) devices. In networks with multiple devices and clouds, factors such as real-time computing requirements, fluctuating wireless channel conditions, and dynamic network scales introduce significant complexity. Addressing these challenges, along with the resource constraints of IoT devices, is essential for effective multi-cloud integration.  This paper proposes a hybrid decision-offloading model that integrates continuous and discrete decision-making. IoT devices must learn to make coordinated decisions regarding cloud server selection, task offloading ratios, and local computation capacity. This dual-layer decision-making process involves managing both continuous and discrete variables, along with inter-device coordination, which poses considerable challenges. To address these, we introduce a probabilistic approach that transforms discrete actions, such as selecting a cloud server, into a continuous domain. We further develop a Privacy-Aware Multi-Agent Deep Reinforcement Learning (PA-MADRL) framework that combines centralized training with distributed execution. This framework minimizes overall system costs by considering energy consumption and cloud server rental fees. Each IoT device operates as an agent, autonomously learning efficient policies while alleviating its computational burden.  Experimental results demonstrate that the PA-MADRL framework effectively adapts to dynamic network conditions, learning optimal offloading policies. It significantly outperforms four state-of-the-art deep reinforcement-learning models and two heuristic methods, achieving lower system costs and improved resource efficiency. 2025 2025 47 60 10.54216/JCIM.160204 https://www.americaspg.com/articleinfo/2/show/3771