Significant b reakthroughs in machine learning (ML) and the advantages of on-device processing have led to edge devices increasingly incorporating accelerators like GPUs, NPUs, and DSPs. However, these accelerators consume energy, prompting users to limit their floating-point precision. Many edge device users are in regions where including high-fidelity accelerators is too costly, leading to low-cost devices with low precision, sacrificing accuracy. Previous work predetermined layer assignments between the CPU and accelerator offline for high accuracy and low latency without considering the input, but we observe that input affects optimal layer assignment. To address this, we present Flex, a system for Fast, Accurate DNN Inference on Low-Cost Edges using Heterogeneous Accelerator eXecution. Leveraging common observations from models on various edge devices, Flex uses a lightweight heuristic and reinforcement learning (RL) to dynamically assign layers across the CPU and accelerator. Experiments show Flex improves average inference time by up to 39%, accuracy by up to 22%, and energy consumption by up to 61% compared to state-of-the-art methods, and is only 4.2% less optimal than the best achievable results.

Air travel often lacks cohesive digital tools to support both pre-trip organization and engaging in-flight experiences. We present FlyEnJoy, a mobile application with an offline-first architecture that integrates itinerary planning management, and in-flight social features to improve travel satisfaction. The system highlights local peer-to-peer networking for realtime in-flight interaction without the Internet and context-aware smart planning tools that adapt to each trip. This demo paper outlines the design and technical components of FlyEnJoy, including its offline-first design, peer-to-peer networking and context-aware smart planning approach. We also discuss insights from a traveler survey that motivated the need for such a platform and conclude with a demonstration that showcases the capabilities of FlyEnJoy in a live scenario.

Connected and autonomous vehicles (CAVs) will be equipped with onboard deep neural network (DNN) models for processing the data from different sensors and communication units in CAVs and on the roads. In the CAV scenario, each vehicle receives time-series driving signals (e.g., speed, brake status) from nearby vehicles, uses its onboard DNN model for driving maneuver prediction for the nearby vehicles, and then takes corresponding actions for safe and efficient driving. Several black-box adversarial attacks for DNN maneuver classifiers in the CAV scenario have been proposed, in which an attacker deliberately sends false driving signals (i.e., add minimum perturbation to the predicted input) to its nearby vehicle to fool its onboard DNN model to output a wrong class label and cause unwanted traffic incidents or congestion. Though previous research proposed adversarial attack detection methods, the methods are for general DNN models rather than specifically for the time-series driving maneuver classification. To detect such adversarial attacks on the DNN maneuver classifiers in the CAV scenario, in this paper, we analyze the adversarial attacks and propose four different approaches comprised of both statistical and machine learning (ML) based methods. Our trace-driven and real experiments show the combination of all four approaches performs 14% better in accuracy rate and 12% better in precision rate compared to the state-of-the-art statistical and ML-based methods. These improvements are critical for driving safety in CAV scenarios.

With the emergence of edge computing along with its local computation advantage over the cloud, methods for distributed deep learning (DL) training on edge nodes have been proposed. The increasing scale of DL models and large training dataset poses a challenge to run such jobs in one edge node due to resource constraints. However, the proposed methods either run the entire model in one edge node, collect all training data into one edge node, or still involve the remote cloud. To handle the challenge, we propose a fully distributed training system that realizes both Data and Model Parallelism over a network of edge devices (called DMP). It clusters the edge nodes to build a training structure by taking advantage of the feature that distributed edge nodes sense data for training. For each cluster, we propose a heuristic and a Reinforcement Learning (RL) based algorithm to handle the problem of how to partition a DL model and assign the partitions to edge nodes for model parallelism to minimize the overall training time. Taking advantage of the feature that geographically close edge nodes sense similar data, we further propose two schemes to avoid transferring duplicated data to the first-layer edge node as training data without compromising accuracy. Our container-based emulation and real edge node experiments show that our systems reduce up to 44% training time while maintaining the accuracy comparing with the state-of-the-art approaches. We also open sourced our source code.

With the emergence of edge devices along with their local computation advantage over the cloud, distributed deep learning (DL) training on edge nodes becomes promising. In such a method, the cluster head of a cluster of edge nodes schedules all the DL training jobs from the cluster nodes. Using such a centralized scheduling method, the cluster head knows all the loads of the cluster nodes, which can avoid overloading the cluster nodes, but the head itself may become overloaded. To handle this problem, we first propose a multi-agent RL (MARL) system that enables each edge node to schedule its own jobs using RL. However, without the coordination between the nodes, action collision may occur, in which multiple nodes may schedule tasks to the same node and make it overloaded. To avoid these problems, we propose a system called Shielded ReinfOrcement learning (RL) based DL training on Edges (SROLE). In SROLE, each edge node schedules its own jobs using multi-agent RL. The shield deployed in a node checks action collisions and provides alternative actions to avoid the collisions. As the central shield node for the entire cluster may become a bottleneck, we further propose a decentralized shielding method, in which different shields are responsible for different regions in the cluster and they coordinate to avoid action collisions on the region boundaries. Our container-based emulation experiments show that SROLE reduces training time by up to 59% with 29% lower median resource utilization and reduces the number of action collisions by up to 48% compared to multi-agent RL and the centralized RL. Our real device experiments show that SROLE still reduces the training time by up to 53% with 28% lower median resource utilization than multi-agent RL and the centralized RL.

Recent research has demonstrated that machine learning algorithms are vulnerable to adversarial attacks, in which small but carefully crafted input perturbations can lead to algorithm failure. It has been demonstrated that certain adversarial attack algorithms are capable of producing these types of perturbations. These attack methods are inapplicable when the attack must be generated in near real time. The use of a hardware accelerator, such as a Process-in-Memory (PIM) archi-tecture, is a potential method for addressing this issue. The PIM architecture is regarded as a superior option for data-intensive applications such as solving optimization problems and Deep Neural Networks (DNN) due to its capacity for ultra-low-latency parallel processing. However, implementing an adversarial attack algorithm directly on the PIM platform is inefficient due to the PIM architecture's complexity and overhead costs. To address this issue, we utilize a novel adversarial attack scheme based on the PIM that leverages Look-up-Table (LUT)-based processing. The proposed LUT-based PIM architecture is capable of being dynamically programmed to execute the operations necessary for an adversarial attack algorithm. Our simulations reveal that the proposed method is capable of achieving an ultra-low operating delay and energy-efficiency performance.

Even though memory denial-of-service attacks can cause severe performance degradations onco-locatedvirtual machines, a previous detection scheme against such attacks cannot accurately detect the attacks and also generates high detection delay and high performance overhead since it assumes that cache-related statistics of an application follow the same probability distribution at all times, which may not be true for all types of applications. In this paper, we present the experimental results showing the impacts of memory DoS attacks on different types of cloud-based applications. Based on these results, we propose two lightweight and responsive Statistical based Detection Schemes (SDS/B and SDS/P) that can detect such attacks accurately. SDS/B constructs a profile of normal range of cache-related statistics for all applications and use statistical methods to infer an attack when the real-time collected statistics exceed this normal range, while SDS/P exploits the increased periods of access patterns for periodic applications to infer an attack. Upon SDS, we further leverage deep neural network (DNN) techniques to design a DNN-based detection scheme that is general to various types of applications and more robust to adaptive attack scenarios. Our evaluation results show that SDS/B, SDS/P and DNN outperform the state-of-the-art detection scheme, e.g., with 65% higher specificity, 40% shorter detection delay, and 7% less performance overhead. We also discuss how to use SDS and DNN-based detection schemes under different situations.

The current autonomous vehicles are equipped with onboard deep neural network (DNN) models to process the data from different sensor and communication units. In the connected autonomous vehicle (CAV) scenario, each vehicle receives time-series driving signals (e.g., speed, brake status) from nearby vehicles through the wireless communication technologies. In the CAV scenario, several black-box adversarial attacks have been proposed, in which an attacker deliberately sends false driving signals to its nearby vehicle to fool its onboard DNN model and cause unwanted traffic incidents. However, the previously proposed black-box adversarial attack can be easily detected. To handle this problem, in this paper, we propose a Suspicion-free Boundary Black-box Adversarial (SBBA) attack, where the attacker utilizes the DNN model's output to design the adversarial perturbation. First, we formulate the attack design problem as a goal satisfying optimization problem with constraints so that the proposed attack will not be easily detectable by detection methods. Second, we solve the proposed optimization problem using the Bayesian optimization method. In our Bayesian optimization framework, we use the Gaussian process to model the posterior distribution of the DNN model, and we use the knowledge gradient function to choose the next sample point. We devise a gradient estimation technique for the knowledge gradient method to reduce the solution searching time. Finally, we conduct extensive experimental evaluations using two real driving datasets. The experimental results show that SBBA outperforms the previous adversarial attacks by 56% higher success rate under detection methods, 238% less time to launch the attacks, and 76% less perturbation (to avoid being detected), and 257% fewer queries (to the DNN model to verify the attack success).

Applications running in edge computing system seamlessly collect data, process data and take actions accordingly. In many cases, the applications need to assist people in real time, and even have life-or-death consequences such as heart attack detection in healthcare and object detection in driving, which requires low job latency. Moreover, power and bandwidth are constrained resources in edge computing systems. Therefore, a challenge is how to handle data efficiently to reduce job latency, and meanwhile reduce power and bandwidth consumption. Previous works mainly focus on where to store collected source data to reduce the communication latency for source data sharing. Noticing that intermediate and final processing results may be shared by many applications, we propose to store intermediate and final results for sharing to avoid the duplicated computation. We also propose data collection that reduces data collection frequency based on context-related factors to achieve an optimal tradeoff between the overhead and decision making accuracy. We further propose data redundancy elimination to reduce the redundant data transmitted between edge and fog nodes. Our combined data operation strategies show significant improvement over the state-of-the-art methods in terms job latency, power and bandwidth consumption for experiments on both simulated and real edge environment.

Deep Neural Network (DNN) models are expected to be widely used in self-driven autonomous vehicles to understand surrounding environments and enhance driving safety. In this paper, we propose a Fast Black-box Adversarial (FBA) attack for time-series DNN models in connected autonomous vehicle (CAV) scenarios. In this attack, an attacker sends false driving signals to a vehicle to misclassify its DNN model (e.g., maintaining speed is misclassified to stopping). Though different black-box adversarial attacks have been proposed previously, they are mainly for image classification, which cannot be directly adopted in the CAV scenarios due to two challenges. First, the attack needs to be generated in near real time. Second, it should not be noticeable based on the driving time-series signals. To handle these two challenges, FBA consists of two steps for the adversarial signal generation: offline and online. First, based on our real data analysis observation that each driving maneuver has maneuver-specific similar patterns (in the time-series) regardless of drivers or vehicles, FBA finds the influential input portion for each maneuver as the offline adversarial signal portion. Second, given a benign driving signal input, FBA replaces its influential input portion with the offline adversarial signal portion and smooths the signals, and uses this input as the initial solution to find the optimal perturbation (that leads to successful attack while minimizing the perturbation values) online using the zeroth-order gradient descent method. It significantly reduces the time to find the optimal perturbation since the initial solution is closer to the optimal solution. Our experiments based on real-driving datasets show the effectiveness of FBA in dealing with the two challenges compared with the existing black-box adversarial attacks.

Many outages of cloud computing services are caused by the natural disasters or common causes such as software failure, hardware failure, cyber attacks, and power outage. As a result, it is critical to develop resilient and robust cyber-physical system (CPS) to support continuity of service for smart and connected communities (S&CC). In spite of considerate research efforts on virtual machine (VM) migration for enhancing failure-resilience of datacenters, one issue still needs to be effectively addressed: how to determine the best destination for VM migration to avoid the influence from a given failure and enable edge nodes to connect to the VMs continuously. In this paper, we aim to handle these important issues to build a resilient and robust edge-cloud (or fog) network system supporting CPS for S&CC. We propose a machine learning based method for VM migration destination determination for the VMs in the predicted failure domains. We also propose a continuous edge-cloud connection method in wireless network component that enables edge nodes to continuously connect with the VMs through intermediate edge nodes when they cannot directly connect to their original VMs due to VM migration. Our experimental results show our proposed system reduces the number of job failures by 1.8 times and reduces the total job completion time by 48% for completed jobs compared to the case without our system.

In a connected autonomous vehicle (CAV) scenario, each vehicle utilizes an onboard deep neural network (DNN) model to understand its received time-series driving signals (e.g., speed, brake status) from its nearby vehicles, and then takes necessary actions to increase traffic safety and roadway efficiency. In the scenario, it is plausible that an attacker may launch an adversarial attack, in which the attacker adds unnoticeable perturbation to the actual driving signals to fool the DNN model inside a victim vehicle to output a misclassified class to cause traffic congestion and/or accidents. Such an attack must be generated in near real-time and the adversarial maneuver must be consistent with the current traffic context. However, previously proposed adversarial attacks fail to meet these requirements. To handle these challenges, in this paper, we propose a Context- aware Black-box Adversarial Attack (CBAA) for time-series DNN models in CAV scenarios. By analyzing real driving datasets, we observe that specific driving signals at certain time points have a higher impact on the DNN output. These influential spatio-temporal factors differ in different traffic contexts (a combination of different traffic factors (e.g., congestion, slope, and curvature)). Thus, CBAA first generates the perturbation only on the influential spatio-temporal signals for each context offline. In generating an attack online, CBAA uses the offline perturbation for the current context to start searching the minimum perturbation using the zeroth-order gradient descent method that will lead to the misclassification. Limiting the spatio-temporal searching scope with the constraint of context greatly expedites finding the final perturbation. Our extensive experimental studies using two different real driving datasets show that CBAA requires 43% fewer queries (to the DNN model to verify the attack success) and 53% less time than existing adversarial attacks.

Many applications such as healthcare and transportation on edge devices will use deep neural network (DNN) prediction based on time-series data collected by the devices. However, the existence of noises in the on-device sensors negatively impacts the sensing output of the DNN models. The state-of-the-art time-series based DNN approaches can deal with Gaussian noise but cannot effectively handle other types of noises in spite of the existence of different types of noises such as shot, burst, transient noises, and their combination. In this paper, we propose an ensemble-based DNN model, namely E-Sense, which consists of different expert models for different noises and shows higher prediction accuracy. Since an edge device may have limited resources to run a large DNN model, we further propose a novel searching-based model compression method called E-Comp that uses knowledge distillation to compress E-Sense to a smaller DNN model while maintaining the accuracy. Our real experiments on live sensor data and trace-driven experiments on three real traces show that E-Sense outperforms other methods in accuracy, and E-Comp reduces 27% inference time without sacrificing accuracy compared with other DNN compression methods. We also distributed our source code.

The key to optimizing the performance of an anycast-based system (e.g., the root DNS or a CDN) is choosing the right set of sites to announce the anycast prefix. One challenge here is predicting catchments. A naïve approach is to advertise the prefix from all subsets of available sites and choose the best-performing subset, but this does not scale well. We demonstrate that by conducting pairwise experiments between sites peering with tier-1 networks, we can predict the catchments that would result if we announce to any subset of the sites. We prove that our method is effective in a simplified model of BGP, consistent with common BGP routing policies, and evaluate it in a real-world testbed. We then present AnyOpt, a system that predicts anycast catchments. Using AnyOpt, a network operator can find a subset of anycast sites that minimizes client latency without using the naïve approach. In an experiment using 15 sites, each peering with one of six transit providers, AnyOpt predicted site catchments of 15,300 clients with 94.7% accuracy and client RTTs with a mean error of 4.6%. AnyOpt identified a subset of 12 sites, announcing to which lowers the mean RTT to clients by 33ms compared to a greedy approach that enables the same number of sites with the lowest average unicast latency.

In this poster, we present measurement studies of the impact of memory DoS attacks on different types of cloud-based applications. Based on the observations, we propose a lightweight, responsive Statistical based Detection Scheme (SDS) that can detect such attacks accurately. Our initial evaluation results show that SDS outperforms the state-of-the-art detection scheme up to 2% higher recall, up to 65% higher specificity, and up to 40% shorter detection delay.

Even though memory-based denial-of-service attacks can cause severe performance degradations on co-located virtual machines, a previous detection scheme against such attacks cannot accurately detect the attacks and also generates high detection delay and high performance overhead since it assumes that cache-related statistics of an application follow the same probability distribution at all times, which may not be true for all types of applications. In this paper, we present the experimental results showing the impacts of memory DoS attacks on different types of cloud-based applications. Based on these results, we propose two lightweight, responsive Statistical based Detection Schemes (SDS/B and SDS/P) that can detect such attacks accurately. SDS/B constructs a profile of normal range of cache-related statistics for all applications and use statistical methods to infer an attack when the real-time collected statistics exceed this normal range, while SDS/P exploits the increased periods of access patterns for periodic applications to infer an attack. Our evaluation results show that SDS/B and SDS/P outperform the state-of-the-art detection scheme, e.g., with 65% higher specificity, 40% shorter detection delay, and 7% less performance overhead.

Connected and autonomous vehicles (CAV) have been introduced to increase roadway safety and traffic flow efficiency. In CAV scenarios, an autonomous vehicle shares its current and near-future driving maneuvers in terms of different driving signals (e.g., speed, brake pedal pressure) with its nearby vehicles using wireless communication technologies. Deep neural network (DNN) models are usually used to process the driving maneuver time-series data over other machine learning algorithms due to the high prediction accuracy of DNN models. In this scenario, an attacker can send false driving maneuver signals to fool the DNN model to misclassify an input. The existing black-box adversarial attacks (which are for image datasets) require many queries to the DNN model to check if a generated attack will be successful (hence long time) or high amount of perturbation (low imperceptibility), and thus cannot be applied to the time-sensitive CAV scenarios featured by multi-dimensional time series driving data. In this paper, we present an Advanced black-box Adversarial Attack (A3) for the deep driving maneuver classification models. We first formulate an optimization problem for the attack generation with continuous search space to reduce the search time. To solve the optimization problem, A3innovatively combines the binary search and optimization algorithm to improve the time-efficiency of searching the optimal solution. It first uses a binary partition technique to reduce the perturbation search space in solving the problem to improve time-efficiency. It then uses the zeroth-order stochastic gradient descent approach, which is featured by searching a solution faster for high-dimensional datasets, thus further improving time-efficiency. We evaluate the proposed A3attack in terms of different metrics using two real driving datasets. The experimental results show that the A3attack requires up to 84.12% fewer queries and 57.67% less perturbation with 94.87% higher success rates than the existing black-box adversarial attacks.

Edge computing usage in many applications, such as transportation and healthcare, has been becoming popular nowadays. These applications often use deep learning (DL) prediction, which are highly dependent on time-series data collected by the sensors in the edge devices. However, the presence of noise in the on-device sensors negatively affects the sensing output of the DL models. Recently proposed time-series based DL approaches (e.g., SADeepSense) address this issue with the assumption that in the presence of noise, the correlation of sensor inputs in an edge device changes. In this paper, through real experiments, we notice that this assumption may not hold true in the presence of shot noise. To handle this problem, in order to further improve the prediction accuracy, we propose a DL model, namely NoiseSenseDNN, which more accurately extracts the correlation between different sensor inputs over time in the presence of both shot and white noise due to its unique architecture. We further propose a compressed version of NoiseSenseDNN that minimizes the inference time and consumed energy of the edge device while meeting the accuracy requirement. Our experiments on a workstation and a real edge device and three real traces show that NoiseSenseDNN outperforms SADeepSense in accuracy, and the compressed NoiseSenseDNN significantly reduces inference time and energy consumption while meeting the required accuracy.

The proliferation in the use of the Internet of Things (IoT) and Machine Learning (ML) techniques in edge computing systems have paved the way of using Intelligent Cognitive Assistants (ICA) for assisting people in working, learning, transportation, healthcare, and other activities. A challenge here is how to schedule application tasks between the three tiers in the edge computing system (i.e., remote cloud, fog and edge devices) according to several considered factors such as latency, energy, and bandwidth consumption. However, the state-of-the-art approaches for this challenge fall short in providing a schedule in real time for critical ICA tasks due to complex calculation phase. In this paper, we propose a novel ReInforcement Learning based Task Assignment approach, RILTA, that ensures the timeliness guaranteed execution of ICA tasks with high energy efficiency. We first formulate the task-scheduling problem in the edge computing systems considering timeliness and energy consumption in ICA applications. We then propose a heuristic for solving the problem and design the reinforcement model based on the output of the proposed heuristic. Our simulation results show that RILTA can reduce the task processing time and energy consumption with higher timeliness guarantee in comparison to other existing methods by 13 - 22% and 1 - 10% respectively.