Mobile devices face diverse resource constraints and non-IID data class distributions, requiring fast on-device inference for local in-distribution (ID) classes and on-demand remote support for client-specific out-of-distribution (OOD) classes. Hybrid split federated learning (Hybrid SFL) couples personalized client-side front ends (supporting early exit) with a generalized server-side backend for fallback inference, balancing accuracy and cost. However, under client architectural heterogeneity, the existing hybrid SFL suffers from representation skew, where features from customized extractors fail to align in the shared space, leading to a sharp degradation in the server model responsible for OOD prediction. We propose HARMONY, the first hybrid SFL framework to support heterogeneous client architectures. HARMONY modifies meta-learning to simulate diverse extractors across parameters and architectures, and to learn to personalize. To mitigate representation skew, HARMONY conducts server-side contrastive learning to align extracted features, neither sacrificing clients' personalization nor sharing raw labels. Compared to the state of the art across multiple datasets and model families, HARMONY improves test accuracy by up to 43.0

AI agents are enabling a new paradigm of agent-augmented real-time communication (RTC), where humans focus on high-level collaboration, while agents autonomously retrieve, analyze, and generate information in real time to support their interactions. These apps enable new experiences across various domains: for example, when corporate employees co-author a legal document, their agents can discuss and draft on their behalf, sparing them the burden of manually reviewing each other's work. As existing cloud-based agents suffer from privacy risks and unscalable server costs, on-device agent-augmented RTC offers a promising alternative. However, this on-device paradigm introduces a new networking challenge: contention between concurrent traffic flows generated by humans (for live video streaming) and agents (for sending context files for analysis). We design HFS, a framework to ensure both high live video quality and low agent response latency in agent-augmented RTC apps. We achieve the goal through an app-guided multi-flow transport approach, where a unified app-layer orchestrator jointly controls the sending rates of live video and agent context flows based on their heterogeneous app requirements. Our prototype built atop WebRTC and llama.cpp demonstrates that HAFS outperforms baselines, achieving 1.5x higher video quality while reducing agent response time by 31%.

With mobile video calls now ubiquitous, ensuring seamless video-based real-time communications (RTC) remains a critical challenge for 5G operators. Despite abundant 5G bandwidth, video calls frequently experience low quality and unacceptable latency during channel fluctuations—not due to bandwidth limitations or congestion, but because of non-congestive delays in the radio access network (RAN). These delays stem from general-purpose RAN transmission procedures that prioritize radio resource efficiency over application latency through reactive scheduling and static timer-based retransmissions. Existing solutions largely address congestion-induced delays or sacrifice spectral efficiency to mitigate non-congestive delays.To address this, we present PAVE, a novel RAN-side solution that breaks the fundamental tradeoff between spectral efficiency and RTC latency. Our key insight is that RTC traffic exhibits distinct characteristics—periodic traffic generation and deadline-driven urgency—that can be leveraged to optimize RTC quality of experience (QoE) without sacrificing efficiency. For practical deployment, PAVE extracts these characteristics at the RAN, enables resource preallocation that goes beyond strictly periodic traffic patterns, and incorporates a selective retransmission and skip mechanism that maintains high spectral efficiency while strategically leveraging application-layer recovery. Implemented in an Open-RAN compliant RAN Intelligent Controller, PAVE improves tail frame rates by 1.8× and reduces video stalls by 94% in real-world evaluations.

Delay-based congestion control algorithms (CCAs) have been proposed to tackle the bufferbloat problem of traditional loss-based CCAs. However, existing delay-based CCAs either fail to adequately consider the non-congestive delay caused by scheduling characteristics of modern cellular networks leading to improper congestion control, or face practical deployment issues, which results in an inability to fully utilize the high bandwidth and low latency that recent cellular systems provide. To resolve this problem, we propose Cesar, a cellular resource scheduling-aware congestion control with only sender-side modification. Cesar estimates scheduling unit through TCP ACK interval patterns to deduce the scheduling characteristics of the current cellular link, and adjusts the congestion window size based on scheduling unit in a step-wise manner to minimize the impact of the scheduling delay on congestion control. Experimental results on 5G and LTE cellular networks of three different mobile carriers show that Cesar outperforms other state-of-the-art CCAs. Results show that throughput-over-latency performance improves by up to 2.89x, 10.09x, 1.39x, and 5.65x compared to ExLL, PropRate, BBR, and Cubic, respectively.

The advent of 5G networks promises to revolutionize real-time applications with its high bandwidth capabilities. Despite this, current real-time applications such as 360-degree video conferencing and intelligent transport systems, which demand high data rates and stringent low latency, often find it difficult to satisfy their QoS needs with a single 5 G connection due to inconsistent bandwidth. As a viable option, multipath transport solutions using multiple 5G connections have emerged with recent technological advancements. In this paper, we explore real-time application performance using MPTCP over 5G multiple connections. Our real-world measurements over MPTCP across major 5G operators in South Korea reveal average frame delays exceeding 213 ms. This delay is much higher than our expectations compared to measurements over single-path TCP (SPTCP) under similar network conditions. We attribute this to the delayed reaction of the MPTCP scheduler, which relies on long-term measurements rather than adapting to recent network conditions. The delayed reaction causes sub-optimal path selection and increased latency. This paper contributes a thorough analysis of MPTCP performance degradation in real-time applications, identifies the limitations of current MPTCP schedulers, and considers potential solutions to improve their responsiveness to achieve better application latency performance.

Radio Access Network (RAN) slicing is a key technology in 5G communication systems. It dynamically allocates network resources such as bandwidth and time slots to each RAN slice, meeting the quality of service (QoS) requirements of each slice on a common underlying 5G infrastructure. This RAN slicing problem normally has a large number of resource combinations with a practical number of RAN slices. However, most Q -learning based Deep reinforcement learning (DRL) algorithms cannot successfully converge with the size of the action space. To address this issue, we introduce the architecture of Action Factorization (AF) with a soft-max layer, which aids exploration by decomposing a large action space into multiple independent sub-action spaces. In addition, RAN slicing problem is facing a performance issue for highly mobile users. To improve the performance of this problem, we use current channel information and future channel information predicted by long short-term memory (LSTM). We then propose a DRL architecture combined with AF and LSTM for bandwidth allocation in RAN slicing. Furthermore, we point out that the QoS requirements used as a performance metric in existing studies are inconsistent with the QoE achievement from the user's point of view. Therefore, we introduce new metrics, data rate indicators (DRI), to compensate the discrepancy. Through extensive simulations, we confirm that our proposed solution efficiently allocates bandwidth to each slice for a reasonable number of slices by maximizing the sum of rewards from QoE achievement for each user under high mobility.

Mobile users in an ultra-dense millimeter-wave cellular network experience handover events more frequently than in conventional networks, which results in increased service interruption time and performance degradation due to blockages. Multi-connectivity has been proposed to resolve this, and it also extends the coverage of millimeter-wave communications. In this paper, we propose a dual-connection based handover scheme for mobile UEs in an environment where they are connected simultaneously with two millimeter-wave cells to overcome frequent handover problems. This scheme allows a mobile UE to choose its serving link between the two mmWave connections according to the measured SINRs and then the corresponding base stations may forward duplicate packets to the UE. We compare our dual-connection based scheme with a conventional single-connection based scheme through ns-3 simulation. The simulation results show that the proposed scheme significantly reduces handover rate and delay. Therefore, we argue that the dual-connection based scheme helps mobile users achieve performance goals they require in ultra-dense cellular environments.

The 3rd Generation Partnership Project (3GPP) finalized Release 15 specifications for the 5th Generation New Radio (5G NR) in June 2018. In Release 16, the 3GPP worked on not only technical improvements over the previous release but also the introduction of new features. One of the new features is the use of Two-step Random Access Channel (2-step RACH) that enhances 4-step random access with respect to radio resource control connection setup and resume procedures. In this article, we first look into details of 2-step random access defined in the 3GPP Release 16, and briefly introduce recent literature related to 2-step random access. Second, we present challenges of the above random access schemes. Among the challenges, we focus on how a User Equipment (UE) performs self-uplink synchronization with the next-generation Node B (gNB) to resolve preamble collisions, which occur when multiple UEs transmit the same preamble. Specifically, we propose a framework that helps the UE estimate the Timing Advance (TA) command using a deep neural network model and to determine the TA value. Finally, we evaluate the proposed framework in terms of the accuracy of TA command estimation, the inference time, and the battery consumption.

The ultra-dense network (UDN) is a promising technology that overcomes the instability of high-frequency millimeter wave (mmWave) communication in a cellular network. With a short distance between cells and variability of mmWave, a handover occurs frequently in UDN with mmWave. Hence fast handover is essential to provide reliable service. One of the important things for fast handover is prompt signaling between base stations (BSs) through backhaul. However, conventional handovers assumed backhaul has very short latency with ideal deployment. In practice, with non-ideal latency and deployment of backhaul, handover is delayed for backhaul. We find that this delayed handover led to the performance reduction of TCP, the dominant traffic of nowadays. So we propose a ProSCH, the novel handover scheme that operates on a practical backhaul network while guaranteeing TCP performance. ProSCH reduces a handover interruption time by considering the backhaul latency in the handover signaling process and uses a TCP proxy on BS to forward data without loss after handover. Through extensive ns-3 simulation, we show that ProSCH outperforms conventional schemes in terms of handover interruption time, backhaul load, throughput, and delay.

Mobile users in an ultra-dense millimeter-wave cellular network experience handover events more frequently than in conventional networks, which results in increased service interruption time and performance degradation due to blockages. Multi- connectivity has been proposed to resolve this, and it also extends the coverage of millimeter-wave communications. In this paper, we propose a dual-connection based handover scheme for mobile UEs in an environment where they are connected simultaneously with two millimeter-wave cells to overcome frequent handover problems. This scheme allows a mobile UE to choose its serving link between the two mmWave connections according to the measured SINRs and then the corresponding base stations may forward duplicate packets to the UE. We compare our dual-connection based scheme with a conventional single-connection based scheme through ns-3 simulation. The simulation results show that the proposed scheme significantly reduces handover rate and delay. Therefore, we argue that the dual- connection based scheme helps mobile users achieve performance goals they require in ultra-dense cellular environments.

With the rapid growth in mobile data traffic, 5G communication has been emerging. 5G communication utilizes mmWave links owing to its large bandwidth, but the high-frequency spectrum is not as reliable as the sub-6GHz bands. This is critical to TCP traffic because the network can not distinguish between link errors from network congestion. Multi-connectivity is a promising technology that provides link robustness at the network level. Meanwhile, TCP proxy, an intermediate node that accelerates TCP, can improve TCP throughput in high speed communication with mmWave link. Even though TCP performance in mmWave networks improves by adopting multi-connectivity and TCP Proxy, there is no study that considers using these two together. We integrate the two schemes and simulate the operation of TCP proxy on the multi-connectivity enabled mmWave network using ns-3 simulation. From simulation results, we find that performance of TCP proxy depends on the location of proxy placement, backhaul delay, and buffer size.