Abstract:
Limited computing resources of IoT devices incur prohibitive latency in processing input data. This triggers new research opportunities toward task offloading systems. Deploying computing servers at existing base stations may not be sufficient. This requests mobile edge servers mounted on unmanned aerial vehicles (UAVs) that provide on-demand mobile edge computing (MEC) services. This talk presents an overview of recent deep reinforcement learning (DRL) approaches, in particular, a multi-agent DRL method where multiple intelligent UAVs cooperatively determine their computations and communication policies. Technical challenges and research opportunities in developing multi-UAV MEC networks are discussed.

Abstract:
Speech covers general aspects of UAM Service, Market, Ecosystem, and Experimental Demonstrations. R&D results on VSLAM-based UAV Trajectory Planning using Reinforced Swarm Intelligence are also presented.

Abstract:
To meet the growing mobility needs in intra-city transportation, the concept of urban air mobility (UAM) has been proposed in which vertical takeoff and landing (VTOL) aircraft are used to provide a ride-hailing service. UAMs are expected to revolutionize the transportation industry by providing new aerial mobility options that can change the way in which we travel. However, the effective deployment of UAMs is contingent upon the presence of a reliable wireless and machine learning infrastructure that can enable UAM aircraft to communicate and navigate autonomously in dynamic environments. In this talk, we first present a rigorous performance analysis on the connectivity requirements for UAM systems. Then, leveraging these connectivity results, we propose a novel wireless-enabled asynchronous federated learning (AFL) framework that uses a Fourier neural network to tackle the challenging problem of turbulence prediction during UAM operations. We then show various properties of this framework, and we discuss key results related to learning and communications with UAM. We conclude the talk with some open problems in this field as well as in other related non-terrestrial network research areas.

Abstract:
In our daily lives, various kinds of sounds are produced from people themselves, computing devices, and surroundings. These sounds have plenty of information including when, where and what happened and eventually lead to the emergence of a new class of smart mobile applications. Two interesting examples are introduced through this talk. One is to turn any surface into a touch interface by listening to sounds. When a user taps a surface, an impact sound, called a touchsound, is produced. We can capture it using the built-in microphones of mobile devices and pinpoint the tapping location by analyzing the propagation pattern of the touchsound. Another work is to leverage sounds as a side channel for mobile payment services. When using magnetic induction-based communication technologies, such as NFC (Near-Field Communication) and MST (Magnetic Secure Transmission), some parts of smartphones are slightly deformed and emit sounds. We reveal that these sounds contain sensitive information and can introduce a new security threat for mobile device users.

Abstract:
Massive connectivity is a key to the success of the Internet of Things. While mmWave backscatter has great potential, substantial signal attenuation and overwhelming ambient reflections impose significant challenges. We present OmniScatter, a practical mmWave backscatter with an extreme sensitivity of -115 dBm and scalability up to millions of concurrent tags. At the heart of OmniScatter is the new High Definition FMCW (HD-FMCW), which interplays with the tag (FSK) signal to disentangle the ambient reflections from the tag signal in the frequency domain. This essentially offers immunity to ambient reflections and improves the SNR by over five orders of magnitude! OmniScatter offers coordination-free Frequency Division Multiple Access (FDMA) that scales to millions of concurrent tags without coordination, paving a pathway towards the vision of massive IoT.

Abstract:
Modern high-speed railway (HSR) systems offer a speed of more than 250 km/h, making on-board Internet access through track-side cellular base stations extremely challenging. We conduct extensive measurements on commercial HSR trains, and collect a massive 1.79 TB GPS-labeled TCP-LTE dataset covering a total travel distance of 28,800 km. Leveraging the new insights from the measurement, we design, implement, and evaluate POLYCORN, a first-of-its-kind networking system that can significantly boost Internet performance for HSR passengers. The core design of POLYCORN consists of a suite of novel multipath scheduling algorithms that intelligently determine what, when, and how to schedule user traffic over multiple highly fluctuating cellular links between HSR and track-side base stations. POLYCORN is specially designed for HSR environments through a cross-layer and proactive approach. We deploy POLYCORN on the operational LTE gateway of the popular Beijing-Shanghai HSR route at 300 km/h. Real-world experiments demonstrate that POLYCORN outperforms state-of-the-art multipath schedulers by 31% to 242% in goodput, and reduces the delivery time by 45% for instant messaging applications.

Abstract:
The next generation terrestrial digital broadcasting standard, known as ATSC 3.0 has a variety of new features which the first generation system (ATSC 1.0) does not support, responding to the changes of TV watching trends and needs for the efficient cooperation with broadband networks.
The physical layer of ATSC 3.0 provides various operation modes for different robustness and spectral efficiency required differently for each of intended services. Also, the standard is developed as all-IP based, and therefore, it realizes convergence of broadcast and broadband networks more efficiently. Due to the excellence of ATSC 3.0 technology, it was successfully commercialized in South Korea and the U.S., and furthermore, other countries such as India and Brazil are considering ATSC 3.0 as their next generation terrestrial broadcasting standard.
In this talk, I will review the ATSC 3.0 technologies and the status of the next generation broadcast systems for other countries.
Then, I will introduce the new emerging technologies based on the ATSC 3.0 standard - MIMO-LDM, diversity antenna for mobile reception, and channel bonding – which are being developed to accommodate the recent trends of media services, such as future immersive media, 8K UHD, and mobile services.

Abstract:
Since the 5G mobile service kicked off, with a view to the wider band and beamforming, millimeter-wave antennas have been employed for the communication system. In order to make up for rapid path-loss in the Ka-band, the antenna is in the form of an array which ends up with inevitable loss from the PCB and wirings. To cope with this drawback, metamaterial surfaces are devised to enhance the antenna in light of directivity and gain, without resorting to excessive use of beamformer chipsets. The metamaterial antenna technology is extended to the wireless link of moving vehicles such as LEO satellites. The IITP-RF RRC/MiEMI presents examples of millimeter-wave beamforming antennas generating radiated far-fields of high directivity ascribed to the lensing effect of the metasurface whose size is way smaller than the conventional array antenna at the same frequency. Plus, some of them including the RIS are tested with TEXWave5G the measurement equipment of our own making.

Abstract:
The satellite network is a key element of 6G, and it is expected to overcome the limitations of terrestrial communication and provide communication without spatial restrictions such as deserts or airplanes. Among them, the Low-Earth Orbit (LEO) satellite network, which is actively studied nowadays, is more complex and more time-sensitive than the geostationary orbit (GSO) satellite network. In this talk, I will introduce a solution using a D-wave quantum computer, the first commercialized quantum computer. I will explain how to set the constraints and the objective functions of the LEO satellite network. It will be formulated with Quadratic Unconstrained Binary Optimization (QUBO) so that a quantum computer can solve the problem. Also, I will review the basic quantum properties for a quantum computer to solve QUBO.

Abstract:
Recommender systems have been widely advocated as a way of providing suitable recommendation solutions to customers in various fields such as e-commerce, advertising, and social media sites. In recent years, thanks to the highly expressive capability of graph neural networks (GNNs) for effective representation learning of graphs, GNN-based recommender systems have been developed for improving the recommendation accuracy while exhibiting the state-of-the-art performance. In this talk, I first review existing GNN models for top-K recommendation, which generally were developed using implicit feedback by regarding observed user–item interactions as positive relations. More precisely, I describe state-of-the-art GNN-based recommender systems such as NGCF and LightGCN, which apply convolutions to graph domains, while performing layer aggregation to solve the oversmoothing problem. Then, I address some challenges such that existing GNN-based recommender systems overlook the existence of negative feedback due to their ease of modeling. To tackle this challenge, we discuss how to make use of low rating scores for representing users’ preferences since low ratings can still be informative in designing recommender systems. As a practical solution, I present SiReN, a new Sign-aware Recommender system based on GNN models and its performance superiority over state-of-the-art methods through comprehensive experiments.

Abstract:
Oftentimes humans can immediately respond to unseen events without instantaneously communicating with others. This is thanks to the ability of reasoning about future events, others’ minds, and their possible reactions to the events based on the background knowledge accumulated from experiences and communication in the past. In this talk I will trace back to the fundamentals of human cognition and artificial intelligence (AI), and introduce the ways to deal with semantics and knowledge in various desciplines including linguistics, semiotics, logic, and category theory. Inspired from this, I will introduce several promising ways towards designing AI and semantics native 6G communication with selected examples using multi-agent reinforcement learning and probabilistic logic programming.

Abstract:
Communication systems are moving to increasingly higher carrier frequencies, which brings benefits for high-accuracy localization and sensing, but also poses a significant challenge: at these higher carrier frequencies certain hardware impairments (HWIs) will be more pronounced. In this talk, we cover two HWIs (antenna element perturbations and mutual coupling) and show that (i) for communication these HWIs are relatively mild; (ii) for localization and sensing these HWIs lead to significant performance degradations. To mitigate the effect of these degradations, we propose an AI-based approach, learning both new transmit signals and receiver algorithms, based on data over the impaired channel. We show that under this approach, the performance loss can be recovered. Finally, challenges for future work in this area will be discussed.

Abstract:
As deep learning technology becomes advanced, mobile vision applications such as AR/VR are prevalent. Although there exist many studies on the optimization of mobile resource allocation and learning model independently, they cannot reflect realistic mobile environments due to the assumption of fixed distributions for wireless and service request. In this talk, we will discuss the joint optimization of learning model and process/network resources adapting to system dynamics, and development of corresponding algorithm using state-of-the-art Online Convex Optimization (OCO) learning technique. Finally, we show the performance evaluation of the algorithm using recent AI embedded devices.

Abstract:
Years of technological advancements have made it possible for small, portable, electronic devices of today to last for years on battery power, and last forever - when powered by harvesting energy from their surrounding environment. Unfortunately, the prolonged life of these ultra-low-power systems poses a fundamentally new problem. While the devices last for years, programs that run on them become obsolete when the nature of sensory input or the operating conditions change. The effect of continued execution of such an obsolete program can be catastrophic. For example, if a cardiac pacemaker fails to recognize an impending cardiac arrest because the patient has aged or their physiology has changed, these devices will cause more harm than any good. Hence, being able to react, adapt, and evolve is necessary for these systems to guarantee their accuracy and response time. We aimed at devising algorithms, tools, systems, and applications that will enable ultra-low-power, sensor-enabled, computing devices capable of executing complex machine learning algorithms while being powered solely by harvesting energy. Unlike common practices where a fixed classifier runs on a device, we take a fundamentally different approach where a classifier is constructed in a manner that it can adapt and evolve as the sensory input to the system, or the application-specific requirements, such as the time, energy, and memory constraints of the system, change during the extended lifetime of the system.

Abstract:
Artificial intelligence/Machine Learning (AI/ML) is known to be effective in reducing network installation cost, automating network operation, and improving network performance by solving nonlinear problems or complex optimization problems in the mobile communication field.
Recently, there are a lot of researches using AI/ML to design air interface and replacing some existing algorithms which was based on mathematical models in RAN. AI/ML is expected to play a defining role in future mobile communication. This presentation introduces some AI/ML use cases, research results, the current status of 3GPP standardization of AI/ML in the RAN domain.

Abstract:
Autonomous vehicles(AVs),such as drones and self-driving cars, are a type of cyber-physical systemsfor automatedtransportation and missions. With their increasing adoption, AVs are facing threats of cyber and cyber-physical attacks that exploit their attack surfaces. Although many AVs are critical to human safety and the environment, it is difficult to make them secure against such attacks due to new challengesthat are not addressable by traditional approaches. Many of these challenges originate from asemantic gap between cyber and physical domainsin which AVsystems operate.
In this talk, Iwill introducemy recent work that bridgesthegap todiscover securityvulnerabilities in AV systemseffectively.Specifically, I willintroducetwofuzzing toolsthatautomatically explore the input spaces of(1) drone control systemsand (2) self-driving car systems, respectively,and detect hidden vulnerabilities that attackers may exploit tocause critical accidents. I will show how ourfuzzing tools (in the cyber domain)generate sensorinputs and monitor vehicle operations(in the physical domain), and discuss uniquechallenges that we addressed while buildingthesemechanisms.Using our tools, wehave discoveredover 100 new bugs inpopular drone control and self-driving car systems and contributedto the eliminationof the bugsfor the security and safety of the AV systems.

Abstract:
Cross-platform binary analysis requires a common representation of binaries across platforms, on which a specific analysis can be performed. Recent work proposed to learn low-dimensional, numeric vector representations (i.e., embeddings) of disassembled binary code, and perform binary analysis in the embedding space. Unfortunately, however, existing techniques fall short in that they are either (i) specific to a single platform producing embeddings not aligned across platforms, or (ii) not designed to capture the rich contextual information available in a disassembled binary.
In this talk, I will present a deep learning-based method, XBA, which addresses the aforementioned problems. To this end, binaries are first represented as typed graphs, dubbed binary disassembly graphs (BDGs), which encode control-flow and other rich contextual information of different entities found in a disassembled binary, including basic blocks, external functions called, and string literals referenced. Binary code representation learning is then formulated as a graph alignment problem, i.e., finding the node correspondences between BDGs extracted from two binaries compiled for different platforms. XBA uses graph convolutional networks to learn the semantics of each node, (i) using its rich contextual information encoded in the BDG, and (ii) aligning its embeddings across platforms. This formulation allows XBA to learn semantic alignments between two BDGs in a semi-supervised manner, requiring only a limited number of node pairs be aligned across platforms for training. The evaluation results show that XBA can learn semantically-rich embeddings of binaries aligned across platforms without apriori platform-specific knowledge.

Abstract:
The number of everyday smart devices is projected to grow to the billions in the coming decade, which enables various smart building applications. These applications, especially in-home long-term occupant monitoring, rely on the emerging Internet-of-Things sensing techniques. We introduce ‘Structures as Sensors’, where we leverage ambient vibration induced by people to infer their information indirectly and non-intrusively. From the system perspective, general problems faced by sensing technologies, especially for indirect sensing, are the tradeoff between the sensing data efficiency/quality and the system deployment constraints. We address this issue by repurposing human sensing data to learn the information about the deployment. Then we optimize the sensing system to acquire high-quality data for the application accordingly. However, high-quality sensing data may still lead to low prediction accuracy, due to the complexity of the physical world -- sensing data distributions can change significantly under different sensing conditions. Therefore, from the data/learning perspective, accurate information learning through pure data-driven approaches requires a large amount of labeled data, which is costly and difficult to obtain in real-world applications. We address these challenges by combining physical and data-driven knowledge to reduce label data needed via physical knowledge-guided model transfer.