Edge-based distributed deep learning (DDL) is increasingly important for embedded AI scenarios where data resides within a trusted domain (e.g., a hospital or campus network) and compute is pushed to edge nodes for cost or performance reasons. Deploying efficient DDL systems in such settings requires evaluating a large space of training configurations, each defined by a combination of cluster topology and synchronization strategy. Exploring this space through real training is prohibitively slow, and existing single-machine simulators are inaccurate and inefficient, as they require full training runs to estimate timing, overlook communication effects, and suffer timing distortions due to host-level scheduling.

We show that many popular mobile services suffer from excessive network bandwidth consumption. The root cause is that the existing mobile/cloud communication interfaces are designed and optimized for service providers rather than end user devices. Solving the problem is challenging, because of the conflicted interests of service providers and mobile devices. We propose Edge-hosted Personal Service (EPS), with which device-oriented solutions can be easily deployed without affecting service providers. EPS also enjoys other notable advantages, including enabling new mobile services, reducing loads on the cloud, and benefiting delay-sensitive applications. We demonstrate the usefulness of EPS by designing ETA (Edge-based web Traffic Adaptation), an effective solution for the excessive bandwidth consumption problem, and deploy ETA with a prototype EPS system. By exploring lightweight virtualization techniques, our EPS prototype system is highly scalable in terms of concurrent EPS instances, and secure in terms of resource isolation. The real-world evaluation shows that our ETA EPS can effectively reduce bandwidth for mobile devices with small overheads.

Being powered by battery and the reliance on expensive cellular data while users are on the move are holding smart mobile devices from being widely used in those long-lasted, computation-intensive, or highly network-reliant usage scenarios. In the meantime, more and more mobile workloads and optimizations now rely on the cloud, such as mobile cloud offloading, cloud-based mobile web optimization, and cloud-based network traffic redundancy reduction. However, it is hard to perform large-scale and personalized support or optimization for mobile workloads without significant computation resource increase on the cloud, as well as higher bandwidth requirement on the networks. We target solving the above two problems by enabling edge-hosed personalized services (EPS for short). The idea of EPS is twofold. One is enabling developer-customizable and power-and-traffic-efficient network communication on the "last-hop" communication between mobile devices and the network edge. The other is distributing cloud services for mobile workloads to network edge, so that they can be done on a personalized basis, while enjoying much lower communication latency to/from mobile devices.

Smartphone device inputs, such as inputs from touchscreen, sensors, and GPS, carry sensitive user information, but are vulnerable to passive and active attacks. We present our ongoing design and implementation of SDIF, a Secure Device Input Framework for smartphones. The core components of SDIF are a small and dedicated bare-metal hypervisor built using ARM hardware virtualization support and a user-space sandbox framework. They collectively ensures SDIF's support for unmodified OSes and apps. SDIF guarantees the secure device inputs from hardware to target applications with two key designs. The first is a design of a secure path for input data to be delivered from input device drivers to target applications. The second is a design of trusted device data reading, which ensures input data generated by hardware is faithfully put to device driver memory.