Candidate: Xue Qin
Date: June 29, 2026
Time: 10:00 AM
Location: Online
Supervisor: Xuemin (Sherman) Shen
All are welcome!
Abstract:
The Internet of Things (IoT) enables massive numbers of devices to exchange data with each other, edge servers, and cloud platforms to support intelligent, delay-sensitive, and resource-demanding services. However, IoT devices are often constrained by limited energy and computing capabilities. Mobile edge computing (MEC) can address these limitations by bringing computing resources closer to IoT devices, allowing computation-intensive tasks to be offloaded to nearby edge servers. The challenges are that offloaded task data may be exposed to passive eavesdroppers, while offloading decisions must be jointly optimized with communication, computation, and security resource allocation under dynamic wireless channels, stochastic task requirements, and heterogeneous edge resources. In addition, IoT applications have diverse Quality of Service (QoS) requirements, including latency, reliability, energy efficiency, and task accuracy. These challenges become more significant as MEC evolves from single-server settings to collaborative multi-server and large-scale intelligent edge networks.
In this thesis, we investigate secure and efficient computation offloading in dynamic MEC networks. Particularly, we develop a series of solutions that progressively address different system scales, security requirements, and application scenarios, moving from physical-layer secure offloading to collaborative secure edge computing and further to semantic-level privacy-preserving offloading.
First, a secure single-user single-server MEC system is studied for supporting IoT applications with stringent latency and reliability requirements. Partial computation offloading is considered, where each task can be divided into locally processed and remotely offloaded parts. To protect the offloaded data against eavesdropping, physical layer security (PLS) is incorporated through friendly jamming, which degrades the eavesdropper’s channel quality. The joint optimization of the offloading ratio, jammer selection, and computing resource allocation is formulated as a Markov decision process (MDP) with hybrid action spaces. A hybrid-action deep deterministic policy gradient (HDDPG) algorithm is developed to handle both discrete and continuous decisions. The proposed approach can enhance task completion performance while reducing energy consumption under dynamic wireless and task conditions.
Second, the thesis extends the system to a single-user multi-server collaborative MEC scenario to improve resource utilization and support flexible offloading. In this setting, a task can be partitioned and offloaded concurrently to multiple edge servers, allowing the user to exploit distributed computing resources at the network edge. This collaborative model introduces more complex decisions, including task partitioning, multi-server resource allocation, and security control. To address these challenges, an Omni-DDPG framework is proposed to jointly optimize these decisions in a unified learning-based manner. Simulation results demonstrate its advantages in task completion, resource utilization, and energy efficiency, while extended multi-user multi-server simulations further show its scalability.
Third, we investigate the limitations of physical-layer security in large-scale MEC networks. Although friendly jamming can enhance transmission security, it often relies on accurate channel state information, consumes additional jamming resources, and tightly couples security protection with wireless resource allocation. Moreover, for intelligent IoT applications involving structured data, protecting transmitted bits may be insufficient because sensitive semantic information can still be inferred from transmitted representations. To overcome these limitations, a diffusion-based semantic encoding framework is proposed for privacy-preserving edge offloading. Instead of transmitting raw data or conventional encoded features, the proposed method transforms task data into approximately noise-like latent representations before transmission, reducing the risk of semantic information leakage. The edge server then recovers task-relevant features through a reverse diffusion process to support efficient task execution. This framework reduces reliance on external jamming resources and achieves a favorable trade-off among task accuracy, communication efficiency, and privacy protection.
In summary, we have developed a progressive framework for secure computation offloading in MEC networks by jointly considering task offloading, resource allocation, security protection, and privacy preservation across different system scales and application requirements. Our solutions provide algorithmic foundations and practical insight for building scalable, secure, and intelligent edge computing systems.