The recent advances of deep learning in various mobile and Internet-of-Things applications, coupled with the emergence of edge computing, have led to a strong trend of performing deep learning inference on the edge servers located physically close to the end devices. This trend presents the challenge of how to meet the quality-of-service requirements of inference tasks at the resource-constrained network edge, especially under variable or even bursty inference workloads. Solutions to this challenge have not yet been reported in the related literature. In the present paper, we tackle this challenge by means of workload-adaptive inference request scheduling: in different workload states, via adaptive inference request scheduling policies, different models with diverse model sizes can play different roles to maintain high-quality inference services. To implement this idea, we propose a request scheduling framework for general-purpose edge inference serving systems. Theoretically, we prove that, in our framework, the problem of optimizing the inference request scheduling policies can be formulated as a Markov decision process (MDP). To tackle such an MDP, we use reinforcement learning and propose a policy optimization approach. Through extensive experiments, we empirically demonstrate the effectiveness of our framework in the challenging practical case where the MDP is partially observable.

Lightweight virtualization technology represented by the container is developing rapidly and plays an increasingly important role in cloud computing. Container technology is efficient and flexible, but the kernel-sharing property also means its isolation is incomplete, making multi-tenancy container clouds vulnerable to co-resident attacks. The existing placement strategy either only considers the resource allocation problem or only considers the co-resident security problem of the virtualization technology on the same host. In this paper, we present SecCPS, a Secure Container Placement Strategy using deep reinforcement learning in a typical cloud scene. SecCPS learns to find the optimal placement to mitigate co-resident threats between containers running on virtual machines. In the design of SecCPS, we (1) present the challenges of securely placing containers running on virtual machines and establish metrics to describe co-residence in container clouds; (2) develop a secure container placement strategy using deep reinforcement learning as a preventive measure for co-residence; (3) implement the strategy and evaluate its effectiveness and scalability. The results show that SecCPS can effectively reduce co-residence compared with existing strategies and is scalable and flexible to large-scale container deployment scenarios.

With the capability of trading accuracy for latency on-the-fly, the technique of adaptive early-exit inference has emerged as a promising line of research to accelerate the deep learning inference. However, studies in this line of research commonly use a group of thresholds to control the accuracy-latency trade-off, where a thorough and general methodology on how to determine these thresholds has not been conducted yet, especially with regard to the common requirements of average inference latency. To address this issue and enable latency-aware adaptive early-exit inference, in the present paper, we approximately formulate the threshold determination problem of finding the accuracy-maximum threshold setting that meets a given average latency requirement, and then propose a threshold determination method to tackle our formulated non-convex problem. Theoretically, we prove that, for certain parameter settings, our method finds an approximate stationary point of the formulated problem. Empirically, on top of various models across multiple datasets (CIFAR-10, CIFAR-100, ImageNet and two time-series datasets), we show that our method can well handle the average latency requirements, and consistently finds good threshold settings in negligible time.

Bring-Your-Own-Device (BYOD) has been widely viewed as a definite trend among enterprises in which employees bring and use their personal smartphones for work. Despite the perceived opportunities of increasing productivity and reducing costs, BYOD raises severe security and privacy concerns: the corporate networks and data are directly exposed to malware apps running on the personal smartphones. This highlights the necessity for performing real-time mobile malware detection in BYOD environments. Deep learning seems to be a natural choice to handle such detection, due to its state-of-the-art detection effectiveness. However, deep learning inference is usually too computationally complex for resource-constrained smartphones, and the communication overhead of cloud-based inference may be unacceptable. As a result, it is hard to seek the tradeoff between the real-time demand and optimality of detection accuracy. In this paper, we tackle this issue by proposing an endedge coordinated inference approach that can support highlyaccurate and average latency guaranteed malware detection. Our proposed approach integrates the early-exit and model partitioning methods to allow fast, correct and smartphonelocalized inference to occur frequently. Extensive evaluations are carried out, demonstrating that our proposed approach offers a good compromise between detection accuracy and efficiency.

Inference based on deep learning models is usually implemented by exposing sensitive user data to the outside models, which of course gives rise to acute privacy concerns. To deal with these concerns, Donget al.recently proposed an approach, namely the dropping-activation-outputs (DAO) first layer. This approach was claimed to be a non-invertible transformation, such that the privacy of user data could not be compromised. However, In this paper, we prove that the DAO first layer, in fact, can generally be inverted, and hence fails to preserve privacy. We also provide a countermeasure against the privacy vulnerabilities that we examined.

Mobile edge computing (MEC), as an enabling paradigm allowing mobile services to be performed at the edge of mobile networks, has been commonly expected to facilitate real-time services and offload backhaul traffic. However, the security issues raised by innovations in networks always hesitate us to embrace their benefits, and so MEC does. DDoS attacks originating from massive mobile devices in 5G, especially the vulnerable internet-of-things devices, is an unavoidable threat for MEC services. Unfortunately, each MEC node is hard to maintain service availability in the wake of DDoS attacks, owing to its limited hardware resources for defense. Therefore, we are motivated to study the cooperative defense against DDoS attacks for MEC, where defense resources can be shared between MEC nodes. Specifically, we purpose a global orchestration (GO) mechanism to make full use of freedom degrees of cooperative defense with respect to the number of cooperative participants and the amount of cooperative resource. In the process of designing the GO mechanism, we formulate the optimization problem of GO, and exploit the bucket elimination algorithm to find the optimal solutions. To make the GO mechanism practically usable, we propose a scalable approximation algorithm based on the mini-bucket elimination algorithm. Extensive simulations are carried out to validate the effectiveness of GO.