The increasing demand for high-definition live video streaming services on mobile devices is often hindered by unstable network conditions and limited computational capabilities. To address these issues, we introduce an adaptive video super-resolution (VSR) based mobile live streaming method, MOBLIVE. The core idea of MOBLIVE is to selectively offload regions of video frames that critically influence user perception quality to the server-side VSR model for enhancement. To further improve the video quality, we deploy a predictive model to choose the optimal VSR model for each selected region. Additionally, we employ an adaptive graphics processing unit (GPU) scheduling strategy that optimizes the allocation of multiple VSR tasks across multiple GPUs. Experimental results show that our approach outperforms the state-of-the-art method in video multi-method assessment fusion (VMAF) and reduces the latency by an average of 73.7% in the typical networking environment.

ARM TrustZone has become a cornerstone of security in mobile edge devices, but its protective measures often come at the expense of energy efficiency and system performance. Existing optimization methods rely heavily on cloud-based deep learning models and real-time measurements, making them vulnerable to fluctuating workloads and variable network conditions. This paper introduces a novel offline tuning framework for optimizing Trusted Applications (TAs) within the Open Portable Trusted Execution Environment (OP-TEE), an open-source TEE built on ARM TrustZone. Unlike traditional approaches, our method dispenses with on-device measurements and repeated TA compilations. Instead, it uses a predictive model trained on runtime characteristics observed in the Rich Execution Environment (REE) to estimate energy and performance metrics. Guided by these predictions, the framework then applies targeted optimizations through an automated tuning mechanism. Experimental results show that this offline approach accelerates the tuning process by 196× compared to the default method, while delivering a 27.23% performance improvement and a 25% reduction in energy consumption within just 170 s. These gains underscore the practicality and effectiveness of the proposed framework, paving the way for more efficient and adaptive TA optimization.