International Conference on Communication Systems and Networks : [proceedings]

Resource allocation is a fundamental problem in telecom and various other fields. In this work, we consider the issue of scheduling users for efficient radio resource management. Existing scheduling methods focus on maximizing the QoS of users. However, in future generation networks, it is imperative to schedule based on the requirements set forth by application owners. This paper presents a novel approach to time scheduling in 5G networks using model-free Reinforcement Learning (RL). The proposed framework leverages recent advancements in RL-based scheduling to dynamically allocate resources based on real-time network conditions and user-defined requirements (as defined by intents). To reduce the complexity of the action space, we introduce a novel greedy mapping function to convert the small action space of the RL agent into time-scheduling weights. Simulation results show that our approach effectively meets performance goals for throughput and latency, while reducing computational costs by 50% compared to existing methods.

In this article, we propose a Deep Q-Learning-based approach to improve the throughput of a Wireless Powered Communication Network (WPCN) by controlling the transmissions of energy harvesting devices through adaptive beam steering. We examine the dynamics of the slotted ALOHA protocol in conjunction with the charging process of energy harvesting devices to develop a framework for maximising throughput through beam steering, despite partial observability constraints, such as unknown device locations and current charge levels. We design a Deep Q-Network that learns the optimal beam steering policy only from the past observations of transmission attempts by the energy harvesting devices. Extensive simulations illustrate that our proposed approach offers as much as 26% improvement in throughput over deterministic round robin beam steering and 9% over random beam steering strategies for a dense network size of 50 devices. The performance is found to improve with increasing length of past observations.

The Internet of Things (IoT) demands lightweight and secure authentication schemes for resource-constrained multihop networks, where IoT nodes communicate with the gateway through multiple intermediary nodes. Present distributed schemes enable secure lightweight authentication for such networks, but overwhelm an IoT node with multiple authentication requests from its newly joined child nodes, leaving other nodes underutilized. This paper presents a decoupled distributed authentication scheme, where IoT nodes capable of authentication are selected by the gateway as an authentication server to authenticate newly joined nodes. The authentication server may be a parent or a non-parent node and thus separates the process of authentication and node joining, mitigating the skewed authentication load by newly joined child nodes to their parent node. After a successful authentication, the authentication server informs the gateway about the authenticated node and mediates the session key generation between the gateway and the node. It utilizes Physically Unclonable Function (PUF) responses to eliminate the need for long-term storage of the secret key (for the secure authentication execution) and prevents adversarial prediction by introducing session-based randomness. Lightweight hash-based accumulators and XOR operations are also introduced, producing reduced computational overhead on resource-constrained nodes. Security validation of the scheme is performed, both formally (using ProVerif) and informally. In addition, performance analysis is conducted through theoretical and simulation-based frameworks, focusing on energy consumption, CPU time, and communication overhead. Analysis demonstrates that the proposed scheme is secure while maintaining a lightweight structure over existing methods.

WiFi-based indoor localization provides a practical alternative to vision and sensor-based tracking, but its adoption is limited by high energy costs from repeated signal scans. To address this, we propose LowFI, a layout-aware, low-energy WiFi localization framework that jointly optimizes the number of scans and access point (AP) selection. LowFI leverages fading in RF channels to adaptively budget scans and rank APs, reducing redundancy while maintaining accuracy. Experiments with various WiFi-enabled devices show that localization error in nLoS conditions decreases from ≈8 m with a single scan to ≈2 m with much limited number of scans, after which accuracy gains saturate. By selecting minimal scans and a modest set of APs, LowFI achieves sub-meter accuracy while cutting energy cost by up to 65%. These results demonstrate that careful scan and AP orchestration can deliver reliable localization without compromising device battery.

In urban 5G Multi-access Edge Computing (MEC), overlapping server coverage and dense, heterogeneous user service demands often lead to uneven request distribution, causing resource underutilization at some edge servers and overloading at others. This imbalance degrades overall system efficiency, making load balancing essential for effective resource utilization and Quality of Service (QoS) assurance during multiple service placement and request allocation. We address this by formulating the Load-aware Service placement and Request allocation (LSR) problem as a multi-objective optimization model and proposing a scalable and efficient two-stage heuristic. The objective is to minimize server utilization variance while reducing service deployments and maximizing user coverage. In the heuristic, the first stage ensures maximum coverage through deterministic request allocation and greedy service placement, all within MEC constraints. The second stage iteratively refines service placement and request allocation to minimize utilization variance, offering reduced computational complexity compared to existing approaches. Experimental results demonstrate that our approach achieves significantly lower variance (65%) in optimal scenarios, reduces service deployments by 15%, increases user coverage by 15%, and establishes itself as an efficient centralized solution for quasi-static multi-service edge deployments in urban 5G-MEC environments.

Federated Learning (FL) enables decentralized model training without sharing raw data, but model-weight distortion remains a major challenge in resource-constrained IoT networks. In multi-tier Federated IoT (Fed-IoT) systems, unstable connectivity and adversarial interference can silently alter transmitted parameters, degrading convergence. We propose DPEMAR, a differentially private, error-model-based autonomous repair framework that detects and reconstructs transmission-induced distortions during FL aggregation. DP-EMAR estimates corruption patterns and applies adaptive correction before privacy noise is added, enabling reliable in-network repair without violating confidentiality. By integrating Differential Privacy (DP) with Secure Aggregation (SA), the framework distinguishes DP noise from genuine transmission errors. Experiments on heterogeneous IoT sensor and graph datasets show that DP-EMAR preserves convergence stability and maintains near-baseline performance under communication corruption while ensuring strict (ε,δ)-DP guarantees. The framework enhances robustness, communication efficiency, and trust in privacy-preserving Federated IoT learning.

The continuous increase of UAVs, particularly in swarms, creates significant challenges for security and airspace regulation. Traditional RF fingerprinting methods struggle to detect and classify UAV swarms due to overlapping signals and interference. This study introduces RF-AttenNet, a hybrid deep learning model designed to classify mixed UAV signals by analyzing composite RF spectrograms. RF-AttenNet uses dual attention mechanisms, channel and spatial attention to focus on critical spectral features, enabling the model to effectively separate and identify overlapping UAV signals. We have developed custom composite UAV datasets that simulate real-world swarm interference, incorporating both single and mixed UAV classes. RF-AttenNet achieves state-of-the-art performance, outperforming traditional methods with 95.75% accuracy, 96.41% precision, 95.75% recall, and 95.71% F1-score. RF-AttenNet’s ability to separate concurrent UAV signals highlights its effectiveness in complex swarm detection scenarios. RF-AttenNet is scalable and robust, validated on increasingly complex datasets, including up to 10 UAV types, ensuring its applicability to swarm detection and counter-drone operations. This work advances RF-based UAV detection by addressing the challenges posed by overlapping signals in swarm environments and provides a reliable solution for UAV classification in dense airspace.

Conducting local inference using a compact neural network model on IoT devices helps to improve inference time and reduce dependence on remote cloud servers. However, these compact models must be adapted to the local data from on-device sensors, which often follow a different, possibly unknown, domain or distribution over input data than that used to train a centralized pre-trained model. Directly applying domain adaptation techniques to the compact models leads to poor inference performance. Existing techniques fine-tune a large model to first capture the domain shift in the input data coming from the IoT sensors and then use it to perform model compaction via knowledge distillation. Such a scheme is slow during the domain adaptation phase and is often vulnerable to catastrophic forgetting in which the compact model tends to overfit the local training data. We present a scheme for fast domain adaptation using meta-learning which uses a meta-initialized model to quickly adapt to a target domain while also transferring its knowledge to a compact model via knowledge distillation. Furthermore, we use a federated distillation method which prevents the compact model from overfitting on its local dataset by cooperatively sharing knowledge between other compact models of similar domains. Experiments show that the proposed method achieves faster training times for compact models even of unseen domains while also maintaining acceptable level of generalization ability of the compact models.

Traditionally telecom networks are monolithic in nature and vertically integrated. Single entity takes care of network deployment as well as service delivery. Internet service provider takes care of identity, access, transport, core and backhaul functions. Such monolithic and vertical integrated architecture lack technical innovation and economic competition. Monolithic architectures are also prone to proprietary interface implementations which leads to non-interoperability among communication layers, leading to cost inefficiency and network non-scalability. Whereas open and disaggregated network architecture induces innovation, competition, scalability and cost efficiency. There is multiple industry led efforts to transform telecom networks into open, disaggregated and federated architecture. Open RAN in cellular world and OpenRoaming in Wi-Fi world are few of the noticeable efforts. In India, WANI (Wi-Fi Access Network Interface) is one such disaggregated and federated public Wi-Fi initiative where identity, access, core and backhaul functions are unbundled and offered by independent entities, evolving the ecosystem as a federation. Interoperability among these functions is ensured through a secure exchange of well-defined federation protocols i.e., wani tokens. Disaggregation and federation also add up the network complexity. Enabling roaming in such disaggregated and federated network poses multiple technical challenges. Roaming in telecom network enables a compelling business case for the service providers as well as improve network usability for the end users. In this paper, we present current WANI technical framework, learning from initial deployment, existing technical gaps, ongoing efforts to standardize the architecture and a novel architecture choice for efficient roaming in such network.

Modern financial applications demand architectures that ensure high throughput, low latency, and strong resilience. Traditional CRUD-based systems often encounter transactional bottlenecks and limited fault tolerance, making them less suitable for real-time workloads. In this work, we design and benchmark a Kafka-based event-sourcing architecture that integrates CQRS, schema-registry-backed versioning, Zstandard compression, dynamic Dead Letter Queue (DLQ) handling, and RocksDB-backed local state caching.We evaluate this system against a CRUD-based REST+SQL implementation. To ensure a rigorous comparison, we enforce network symmetry by deploying both the Kafka cluster and the PostgreSQL database on managed cloud infrastructure, eliminating placement bias. Benchmarks were conducted on a dataset of 50,000 banking events (generating over 150,000 persistent state transitions) to measure sustained performance. Our findings show that the Kafka-based approach achieves a sustained throughput of 105.4 events/sec (2.69x over CRUD) while maintaining consistent low-latency reads, validating its suitability for production-grade financial platforms.

Wireless communication systems face challenges like Doppler shifts, multi-path fading, and high mobility. This study explores Orthogonal Time Frequency Space (OTFS) modulation combined with machine learning to improve system performance. OTFS operates in the delay-Doppler domain, making it resilient to channel variations. We implement OTFS modulation to generate datasets with different modulation schemes (BPSK, QPSK, OOK, MSK, QAM, ASK) under various Signal-to-Noise Ratio (SNR) conditions. A hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) model is designed for Automatic Modulation Classification (AMC), achieving high accuracy by recognizing signal patterns. The Isolation Forest algorithm is also applied to detect anomalies, improving data quality. Our results demonstrate the effectiveness of AMC and anomaly detection through detailed metrics and visualizations. This work enhances OTFS-based systems and highlights the role of deep learning in addressing modern wireless communication challenges.

5G/6G networks play a major role in the realization of practical systems for Extended Reality (XR) applications. These networks provide high bandwidth and low latency requirements through a typical enhanced Mobile Broad-Band (eMBB) and Ultra-Reliable Low-Latency Communications (URLLC) slices for XR applications. In this paper, we share the lessons learnt from the implementations: i) Private 5G for XR-assisted indoor facility tour and ii) Public 5G for XR-assisted Medical Emergency Care. We evaluate the performance of XR use cases, present the insights, and propose future work based on implementations.

Sample efficiency and robustness remain significant challenges in applying Reinforcement Learning (RL) to dynamic and heterogeneous telecommunications environments. In this paper, we propose a unifying framework for causal reinforcement learning that integrates counterfactual reasoning and guided causal action selection to improve learning efficiency and policy robustness. Our approach introduces two key techniques: (i) Counterfactual Data Augmentation (CDA), which enriches training data by generating hypothetical outcomes for unobserved actions using structural causal models (SCM), and (ii) Causal Action Selection (CAS), which integrates causal-effect estimation into policy optimization to prioritize high-utility interventions. We validate our approach on 2 standard RL benchmarks as well as a domain-specific Synthetic Slicing Environment simulating Layer 2 scheduling in telecommunications networks. We demonstrate that our framework can reduce the number of environment interactions upto 56% to learn an optimal policy. Our results demonstrate that incorporating causal interventions significantly enhances sample efficiency, stability, and performance of trained RL model compared to conventional RL methods, highlighting the potential for safer and more effective deployment of RL in real-world telecom applications.

This paper analyses the performance of an analog network coding (ANC) based underlay cognitive two-way relay (TWR) network with energy harvesting (EH) relay, in which the secondary users have adaptive multi-rate (MR) selection capability depending on the channel state information (CSI). We consider an ANC-based cognitive TWR network with amplify-and-forward (AF) EH relay, in which the transmission power of secondary nodes is limited by the primary receiver’s interference temperature limit (ITL). In this simultaneous wireless information and power transfer (SWIPT) network, the AF-EH relay adopts power-splitting relaying (PSR) protocol for energy harvesting. In particular, we derived closed-form expression for throughput (τ) of cognitive TWR networks with an adaptive multi-rate selection scheme at secondary nodes. Further, the ergodic sum-rate (R) is analyzed to show the effectiveness of multi-rate transmission scheme. The results establish that multi-rate transmission scheme improves performance significantly over the single-rate transmission in the considered cognitive network scenario. Simulation results validate the mathematical analysis and verify the correctness of obtained expressions.

Music can be considered a form of communicative expression that includes our feelings, identities, and social interactions. It crosses linguistic and cultural barriers. However, in the international music industry, plagiarism still poses a danger to morals, innovation, and legal compliance. In this work, we focus on the structural organization of the versional and plagiarized songs along with the original music tracks across five languages. We propose MUSiNET, a complex network architecture where nodes stand in for music tracks, and the significant value of the cosine similarity score is measured based on the intrinsic (audio) and extrinsic (metadata) features, determining the weighted edges connecting the nodes. We use the HITS algorithm to determine the threshold value of the edge weight to keep the significant edges in the proposed MUSiNET. The network displays the dynamics of the association between the original, plagiarized, and versioned songs using the curated dataset of 5000 tracks from the MUSICOSet dataset. The MUSiNET analysis reveals that plagiarized and versioned tracks frequently emerge as connectors that bridge genres, languages, and cultural boundaries. We observe that the composition of versional and plagiarized songs is not an isolated event in the music development community, but an integral relational process to achieve quick popularity based on the existing audience of the original composition.

Live editing systems in the broadcasting and production industries are increasingly converting to IP based on SMPTE ST2110 and other "Media over IP" standards. In such systems, it is necessary to synchronize with the broadcast station's reference time using PTP (Precision Time Protocol, specified by IEEE 1588). In order to achieve this time synchronization, all network switches along the route must support PTP, i.e., if the network delay jitter becomes large, time synchronization will fail. In this presentation, we propose a method for time synchronization even in networks with large jitter by using adaptive filtering to compute time correction. Furthermore, this method demonstrates that even in networks with significant transmission delay jitter, PTP clock variation can be suppressed, enabling time synchronization via PTP. Consequently, it proves feasible to utilize SMPTE ST2110 for video transmission.

Electric vehicles (EVs) are becoming increasingly attractive for last-mile delivery due to their low operating costs and reduced emissions. However, widespread adoption is hindered by high upfront investments, limited public charging infrastructure, and long charging durations. To address these challenges, we explore a complementary strategy that enables EV fleets to generate revenue through vehicle-to-grid (V2G) energy transactions, while improving charging accessibility through on-the-go solar energy harvesting (SEH). In this work, we formulate a vehicle routing problem for an EV fleet equipped with V2G and SEH capabilities with the objective of reducing the overall delivery costs, while accounting for vehicle load capacities and customer time window constraints. We propose a learning-to-optimize approach (LA) that scales efficiently to problem instances involving hundreds of customer locations and multiple discharge station visits for V2G operations. Using the Solomon benchmark datasets, we compare the performance of the proposed LA with a genetic algorithm (GA). Our results show that LA achieves reasonable solution quality while being 18 times faster than GA, demonstrating its effectiveness for large-scale EV fleet route optimization.

Reliable vehicle-to-everything (V2X) communication is essential for safe driving, coordinated traffic, and intelligent transport services. In practice, non-line-of-sight (NLoS) conditions created by tall buildings, heavy vehicles, and complex inter-sections often disrupt roadside communication, leading to higher latency, packet loss, and reduced reliability. Conventional roadside units (RSUs) are not designed to overcome such blockages, which limits their effectiveness in dense urban environments. This work proposes a RIS-enhanced roadside infrastructure that improves V2X communication under NLoS conditions. Reconfigurable intelligent surfaces (RIS) are programmable panels capable of redirecting wireless signals through controlled reflection. An adaptive optimization algorithm combining beamforming and channel estimation is applied to align reflected signals with receiver requirements. This design creates alternative transmission paths, improving both link stability and transmission quality. Simulation studies compare the proposed RIS-aided framework with traditional RSUs in urban and highway scenarios. Evaluation metrics include packet delivery ratio (PDR), latency, throughput, and bit error rate (BER). Results confirm clear performance gains with RIS, including higher delivery rates, reduced error probability, and lower delay. The study shows that RIS-based roadside infrastructure offers a practical solution for reliable V2X communication, supporting advanced vehicular applications in challenging road environments.

Unmanned aerial vehicles (UAVs) serve as agile network providers for end-to-end wireless communication to ground users, including IoT devices, mobile users, and sensors. While most existing approaches rely on centralized control architectures, such systems are prone to single points of failure, scalability issues, and communication bottlenecks that limit adaptability in dynamic environments. We present a distributed deployment framework for multi-UAV networks, called NIRVANA, which combines message-passing-based cluster formation with nature-inspired resilience mechanisms. In NIRVANA, UAVs exchange local information through message passing to enable decentralized coordination, while resilience is enhanced by maximizing recurrent subnetworks, or network motifs, that have been shown to improve stability and fault tolerance in complex systems. Through a two-phase optimization process, the system jointly maximizes ground coverage and motif-based resilience, ensuring balanced performance under dynamic conditions. Results from extensive simulations demonstrate that NIRVANA achieves 20-36% higher robustness and 4-8% greater user coverage compared to state-of-the-art baseline across varying UAV counts.

Kinesthetic and cutaneous cues are essential for humans to perform dexterous manipulation tasks in daily activities. This research focuses on perception based sampler for enhanced human-robot interaction for haptic based applications. The work presents an end-to-end haptic application with a perception-based adaptive sampling. The application entails a scenario where an operator controls a dexterous robotic gripper, which acts as User Equipment (UE). A custom designed gripper acts as a UE to manipulate a real-world object. During interaction, the contact force (tactile data) from the gripper is rendered as kinesthetic feedback through a haptic device. For kinesthetic-based sampler design, psychophysics-based experiments are conducted to identify the Temporal, spatial deadzone and to study the influence of varying stimulus impacts on kinesthetic perception in the context of perceptually adaptive sampling for teleoperation. The key contributions of this work include the development of spatial and temporal deadzones for tactile-assisted dexterous tasks, the design and comparative analysis of perception-driven sampling methods, and the implementation of a human-in-the-loop control system that enhances the manipulation of virtual objects. Additionally, the application is validated by mapping it onto a 3GPP EDGEAPP-based testbed.