Hyesung Kim

On-device machine learning (ML) enables the training process to exploit a massive amount of user-generated private data samples. To enjoy this benefit, inter-device communication overhead should be minimized. With this end, we propose federated distillation (FD), a distributed model training algorithm whose communication payload size is much smaller than a benchmark scheme, federated learning (FL), particularly when the model size is large. Moreover, user-generated data samples are likely to become non-IID across devices, which commonly degrades the performance compared to the case with an IID dataset. To cope with this, we propose federated augmentation (FAug), where each device collectively trains a generative model, and thereby augments its local data towards yielding an IID dataset. Empirical studies demonstrate that FD with FAug yields around 26x less communication overhead while achieving 95-98% test accuracy compared to FL.

In distributed reinforcement learning, it is common to exchange the experience memory of each agent and thereby collectively train their local models. The experience memory, however, contains all the preceding state observations and their corresponding policies of the host agent, which may violate the privacy of the agent. To avoid this problem, in this work, we propose a privacy-preserving distributed reinforcement learning (RL) framework, termed federated reinforcement distillation (FRD). The key idea is to exchange a proxy experience memory comprising a pre-arranged set of states and time-averaged policies, thereby preserving the privacy of actual experiences. Based on an advantage actor-critic RL architecture, we numerically evaluate the effectiveness of FRD and investigate how the performance of FRD is affected by the proxy memory structure and different memory exchanging rules.

Devices at the edge of wireless networks are the last mile data sources for machine learning (ML). As opposed to traditional ready-made public datasets, these user-generated private datasets reflect the freshest local environments in real time. They are thus indispensable for enabling mission-critical intelligent systems, ranging from fog radio access networks (RANs) to driverless cars and e-Health wearables. This article focuses on how to distill high-quality on-device ML models using fog computing, from such user-generated private data dispersed across wirelessly connected devices. To this end, we introduce communication-efficient and privacy-preserving distributed ML frameworks, termed fog ML (FML), wherein on-device ML models are trained by exchanging model parameters, model outputs, and surrogate data. We then present advanced FML frameworks addressing wireless RAN characteristics, limited on-device resources, and imbalanced data distributions. Our study suggests that the full potential of FML can be reached by co-designing communication and distributed ML operations while accounting for heterogeneous hardware specifications, data characteristics, and user requirements.

By leveraging blockchain, this letter proposes a blockchained federated learning (BlockFL) architecture where local learning model updates are exchanged and verified. This enables on-device machine learning without any centralized training data or coordination by utilizing a consensus mechanism in blockchain. Moreover, we analyze an end-to-end latency model of BlockFL and characterize the optimal block generation rate by considering communication, computation, and consensus delays.

This paper proposes a real-time movement control algorithm for massive unmanned aerial vehicles (UAVs) that provide emergency cellular connections in an urban disaster site. While avoiding the inter-UAV collision under temporal wind dynamics, the proposed algorithm minimizes each UAV's energy consumption per unit downlink rate. By means of a mean-field game theoretic flocking approach, the velocity control of each UAV only requires its own location and channel states. Numerical results validate the performance of the algorithm in terms of the number of collisions and energy consumption per data rate, under a realistic 3GPP UAV channel model.

This paper focuses on cognitive radio (CR) internet- of-things (IoT) networks where spectrum sensors are deployed for IoT CR devices, which do not have enough hardware capability to identify an unoccupied spectrum by themselves. In this sensor- enabled IoT CR network, the CR devices and the sensors are separated. It induces that spectrum occupancies at locations of CR devices and sensors could be different. To handle this difference, we investigate a conditional interference distribution (CID) at the CR device for a given measured interference at the sensor. We can observe a spatial correlation of the aggregate interference distribution through the CID. Reflecting the CID, we devise a cognitive random access scheme which adaptively adjusts transmission probability with respect to the interference measurement of the sensor. Our scheme improves area spectral efficiency (ASE) compared to a conventional ALOHA and an adaptive transmission scheme which attempts to send data when the sensor measurement is lower than an interference threshold.

This paper investigates a cellular edge caching design under an extremely large number of small base stations (SBSs) and users. In this ultra-dense edge caching network (UDCN), SBS-user distances shrink, and each user can request a cached content from multiple SBSs. Unfortunately, the complexity of existing caching controls' mechanisms increases with the number of SBSs, making them inapplicable for solving the fundamental caching problem: How to maximize local caching gain while minimizing the replicated content caching? Furthermore, spatial dynamics of interference is no longer negligible in UDCNs due to the surge in interference. In addition, the caching control should consider temporal dynamics of user demands. To overcome such difficulties, we propose a novel caching algorithm weaving together notions of mean-field game theory and stochastic geometry. These enable our caching algorithm to become independent of the number of SBSs and users, while incorporating spatial interference dynamics as well as temporal dynamics of content popularity and storage constraints. Numerical evaluation validates the fact that the proposed algorithm reduces not only the long run average cost by at least 24% but also the number of replicated content by 56% compared to a popularity-based algorithm.