Secure and Efficient Computation Offloading in Mobile Edge Computing Networks

Abstract

The Internet of Things (IoT) enables large numbers of connected devices to support intelligent, delay-sensitive, and computation-intensive applications. However, IoT devices are typically constrained by limited computing capability, storage capacity, and energy supply. Mobile edge computing (MEC) addresses these limitations by deploying computing resources close to IoT devices, allowing computation-intensive tasks to be processed at nearby edge servers. During wireless offloading, IoT devices may transmit either raw task data or computational workloads for edge processing, or compact task-relevant semantic representations for edge inference. These two offloading paradigms aim to achieve reliable and efficient task processing and accurate and communication-efficient edge inference, respectively.

Achieving these objectives in practical MEC networks involves three key challenges. First, due to the broadcast nature of wireless communications, the transmitted information may be intercepted by passive eavesdroppers, resulting in raw-data leakage or semantic privacy risks. Second, time-varying wireless channels, stochastic task arrivals, and fluctuating edge resources make offloading, transmission, and computing decisions highly coupled. Third, different IoT applications impose diverse quality-of-service (QoS) requirements, including latency, reliability, energy efficiency, deadline-aware task completion, and task accuracy. These requirements are often coupled and may conflict with each other.

Considering these challenges, the objective of this thesis is to develop secure, efficient, and QoS-aware computation offloading solutions for dynamic MEC networks under eavesdropping threats. From a user-centric perspective, three research problems are investigated. The considered system progressively evolves from a single-user single-server scenario to a single-user multi-server collaborative scenario and further to a multi-user multi-server scenario. Meanwhile, the protection objective extends from transmission-level data security to representation-level semantic privacy.

First, secure partial computation offloading is investigated in a single-user single-server MEC system. Each divisible task is partitioned between local processing and edge processing, while friendly jammer-assisted physical-layer security is employed to protect the offloaded task data. The offloading ratio, friendly jammer selection, and edge-server CPU-frequency allocation are jointly optimized under dynamic task and wireless conditions. The resulting long-term optimization problem is formulated as a Markov decision process with a hybrid discrete-continuous action space. To solve this problem, a hybrid-action deep deterministic policy gradient (HDDPG) algorithm is developed. The proposed approach supports secure, energy-efficient, and deadline-aware computation offloading by improving task completion performance while reducing energy consumption.

Second, the system is extended to a single-user multi-server collaborative MEC scenario. A divisible task can be partitioned among local computing and multiple nearby edge servers, enabling parallel processing and collaborative use of distributed edge resources. This scenario introduces higher-dimensional and more strongly coupled decisions, including the task partition ratios, friendly jammer selections, and CPU-frequency allocations across multiple servers. An Omni-DDPG framework is developed to jointly optimize these constrained continuous, discrete, and continuous decisions. The proposed framework improves secure task completion, latency, and energy efficiency through coordinated task partitioning, security control, and edge-resource allocation. Extended multi-user simulations further demonstrate the scalability potential of the learned policy in larger collaborative MEC settings.

Third, semantic privacy-preserving offloading is studied in a multi-user multi-server MEC system. Unlike the first two works, which focus on data-oriented secure offloading, this work considers semantic offloading, where IoT devices transmit compact task-relevant latent representations instead of complete raw task data. Although this approach reduces communication overhead, the transmitted representations may still expose sensitive task-related information to passive eavesdroppers. To address this issue, a diffusion-based semantic encoding framework is developed. A forward diffusion process generates approximately noise-like latent representations before wireless transmission, while an authorized edge server employs a learned reverse diffusion model to recover task-relevant representations for downstream inference. The framework is evaluated under wireless channel noise and quantization and compared with autoencoder- and variational autoencoder-based semantic encoders. The results demonstrate a favorable trade-off among task accuracy, semantic privacy, transmission latency, and energy efficiency.

In summary, this thesis presents a progressive investigation of secure and efficient computation offloading across different MEC system scales and protection levels. The first two works develop learning-based control solutions for data-oriented secure offloading, while the third work extends the protection objective to semantic privacy-preserving offloading. The proposed solutions address task offloading, resource allocation, wireless transmission security, and semantic representation protection, providing algorithmic and practical foundations for scalable, secure, and intelligent edge computing systems.