WebRTC-based Resource Offloading in Smart Home Environments

Web platforms face new demands for emerging applications, which use machine learning models such as pose recognition or object detection. These models require significant computing powers in processing enormous inputs such as images or audios for machine learning computation. These demands are also being generated in smart home appliances based on web platforms. Unfortunately, smart home appliances do not generally have built-in input devices, such as cameras or microphones, due to privacy issues and have limited performance compared to mobile devices. This paper proposes a WebRTC-based resource offloading system for web applications, which allows smart home appliances to use resources of nearby mobile devices as if the resources are their own. We implemented the proposed system, performed experiments on the resource offloading framework, and evaluated the performance using five computation-intensive web applications, which use a machine learning model. Our system was able to run machine learning models, through resource offloading to mobile devices, on smart home appliances without an attached camera, and achieved up to 1.5x speedup, compared to local execution with a camera.