Many times, training a large scale deep learning neural network on a single machine becomes more and more difficult for a complex network model. Distributed training provides an efficient solution, but Byzantine attacks may occur on participating workers. They may be compromised or suffer from hardware failures. If they upload poisonous gradients, the training will become unstable or even converge to a saddle point. In this paper, we propose FABA, a Fast Aggregation algorithm against Byzantine Attacks, which removes the outliers in the uploaded gradients and obtains gradients that are close to the true gradients. We show the convergence of our algorithm. The experiments demonstrate that our algorithm can achieve similar performance to non-Byzantine case and higher efficiency as compared to previous algorithms.

The rise of edge computing gives birth to a spectrum of delay-sensitive applications. Many of these applications build their services atop the functionality that the edge nodes quickly negotiate a unique order on the events received from a massive number of client devices, even under very high event rates. To this end, we propose a protocol, called Nomad, for achieving fast event ordering in edge computing environments. Nomad is designed as a consensus protocol that employs a lease-based approach to take advantage of the locality of the unbalanced workload across the system. It also dynamically adjusts the leadership distribution on the edge nodes based on the recent running history, and relies on a cloud-based arbitrator to resolve contentions. Experiments demonstrate that Nomad outperforms the existing solutions, such as Multi-Paxos, Mencius and E-Paxos, in achieving fast event ordering for large-scale, delay-sensitive edge-cloud applications.

Many edge computing systems rely on virtual machines (VMs) to deliver their services. It is challenging, however, to deploy the virtualization mechanisms on edge computing hardware infrastructures. In this paper, we introduce the engineering and research trends of achieving efficient VM management in edge computing. We elaborate on: 1) the virtualization frameworks for edge computing developed in both the industry and the academia; 2) the virtualization techniques tailored for edge computing; 3) the placement and scheduling algorithms optimized for edge computing; and 4) the research problems in security related to virtualization of edge computing.

We present LAVEA, a system built for edge computing, which offloads computation tasks between clients and edge nodes, collaborates nearby edge nodes, to provide low-latency video analytics at places closer to the users. We have utilized an edge-first design to minimize the response time, and compared various task placement schemes tailed for inter-edge collaboration. Our results reveal that the client-edge configuration has task speedup against local or client-cloud configurations.

In recent years mobile smart devices such as tablets and smartphones have exploded in popularity. We are now in a world of ubiquitous smart devices that people rely on daily and carry everywhere. This is a fundamental shift for computing in two ways. Firstly, users increasingly place unprecedented amounts of sensitive information on these devices, which paints a precarious picture. Secondly, these devices commonly carry many physical world interfaces. In this paper, we propose information leakage malware, specifically designed for mobile devices, which uses covert channels over physical "real-world" media, such as sound or light. This malware is stealthy; able to circumvent current, and even state-of-the-art defenses to enable attacks including privilege escalation, and information leakage. We go on to present a defense mechanism, which balances security with usability to stop these attacks.

Mobile devices are now ubiquitous in the modern world. In this paper, we propose a novel and practical mobile-cloud platform for smart mobile devices. Our platform allows users to run the entire mobile device operating system and arbitrary applications on a cloud-based virtual machine. It has two design fundamentals. First, applications can freely migrate between the user's mobile device and a backend cloud server. We design a file system extension to enable this feature, so users can freely choose to run their applications either in the cloud (for high security guarantees), or on their local mobile device (for better user experience). Second, in order to protect user data on the smart mobile device, we leverage hardware virtualization technology, which isolates the data from the local mobile device operating system. We have implemented a prototype of our platform using off-the-shelf hardware, and performed an extensive evaluation of it. We show that our platform is efficient, practical, and secure.

The smartphone has become an important part of our daily lives. However, the user experience is still far from being optimal. In particular, despite the rapid hardware upgrades, current smartphones often suffer various unpredictable delays during operation, e.g., when launching an app, leading to poor user experience. In this paper, we investigate the behavior of reads and writes in smartphones. We conduct the first large-scale measurement study on the Android I/O delay using the data collected from our Android application running on 2611 devices within nine months. Among other factors, we observe that reads experience up to 626% slowdown when blocked by concurrent writes for certain workloads. Additionally, we show the asymmetry of the slowdown of one I/O type due to another, and elaborate the speedup of concurrent I/Os over serial ones. We use this obtained knowledge to design and implement a system prototype called SmartIO that reduces the application delay by prioritizing reads over writes, and grouping them based on assigned priorities. SmartIO issues I/Os with optimized concurrency parameters. The system is implemented on the Android platform and evaluated extensively on several groups of popular applications. The results show that our system reduces launch delays by up to 37.8%, and run-time delays by up to 29.6%.