This paper proposes an AI framework that leverages integrated sensing and communication (ISAC), aided by the age of sensing (AoS) to ensure the timely location updates of the users for a holographic MIMO (HMIMO)-assisted base station (BS)-enabled wireless network. The AI-driven framework aims to achieve optimized power allocation for efficient beamforming by activating the minimal number of grids from the HMIMO BS for serving the users. An optimization problem is formulated to maximize the sensing utility function, aiming to maximize the communication signal-to-interference-plus-noise ratio (SINRc) of the received signals and beam-pattern gains to improve the sensing SINR of reflected echo signals, which in turn maximizes the achievable rate of users. A novel AI-driven framework is presented to tackle the formulated NP-hard problem that divides it into two problems: a sensing problem and a power allocation problem. The sensing problem is solved by employing a variational autoencoder (VAE)-based mechanism that obtains the sensing information leveraging AoS, which is used for the location update. Subsequently, a deep deterministic policy gradient-based deep reinforcement learning scheme is devised to allocate the desired power by activating the required grids based on the sensing information achieved with the VAE-based mechanism. Simulation results demonstrate the superior performance of the proposed AI framework compared to advantage actor-critic and deep Q-network-based methods, achieving a cumulative average SINRcimprovement of 8.5 dB and 10.27 dB, and a cumulative average achievable rate improvement of 21.59 bps/Hz and 4.22 bps/Hz, respectively. Therefore, our proposed AI-driven framework guarantees efficient power allocation for holographic beamforming through ISAC schemes leveraging AoS.

The rapid evolution towards the sixth-generation (6G) networks demands advanced beamforming techniques to address challenges in dynamic, high-mobility scenarios, such as vehicular communications. Vision-based beam prediction utilizing RGB camera images emerges as a promising solution for accurate and responsive beam selection. However, reliance on visual data introduces unique vulnerabilities, particularly susceptibility to adversarial attacks, thus potentially compromising beam accuracy and overall network reliability. In this paper, we conduct the first systematic exploration of adversarial threats specifically targeting vision-based mmWave beam selection systems. Traditional white-box attacks are impractical in this context because ground-truth beam indices are inaccessible and spatial dynamics are complex. To address this, we propose a novel black-box adversarial attack strategy, termed Spatial Proxy Attack (SPA), which leverages spatial correlations between user positions and beam indices to craft effective perturbations without requiring access to model parameters or labels. To counteract these adversarial vulnerabilities, we formulate an optimization framework aimed at simultaneously enhancing beam selection accuracy under clean conditions and robustness against adversarial perturbations. We introduce a hybrid deep learning architecture integrated with a dedicated Feature Refinement Module (FRM), designed to systematically reshaping irrelevant, noisy and adversarially perturbed visual features. Evaluations using standard backbone models such as ResNet-50 and MobileNetV2 demonstrate that our proposed method significantly improves performance, achieving up to an +21.07% gain in Top-K accuracy under clean conditions and up to a +37.32% increase in Top-1 adversarial robustness compared to different baseline models.

Vision-based beam prediction has emerged as a promising alternative to pilot-driven beamforming in millimeter-wave (mmWave) communications, particularly in highly dynamic 6G environments. By leveraging visual sensing modalities, these systems enable low-latency beam selection without relying on extensive pilot transmissions. However, existing approaches are fundamentally limited by the assumption of fixed user distributions and static codebooks, thus rendering them ineffective in scenarios involving continual user arrivals and evolving beam configurations. In this paper, we propose the first continual learning framework for vision-based beam prediction that adaptively incorporates new user types and expanding codebooks without full model retraining. To ensure scalable adaptation, the proposed method integrates task-driven learning with knowledge distillation and Replay Buffer. As new user types emerge, the model incrementally expands its output space to reflect the updated codebook while retaining a compact Buffer of past samples for stability. A dual-objective optimization governs model updates, thereby combining a supervised loss over new data with a temperature-scaled divergence term that aligns current and prior model predictions on Replayed instances. This mitigates catastrophic forgetting and ensures robust performance across tasks. Experimental results demonstrate that the proposed framework effectively maintains high beam prediction accuracy under continual adaptation, while also achieving lower power loss and enhanced normalized received power relative to baseline methods.

The Open Radio Access Network (O-RAN) paradigm offers a compelling solution to the constraints of traditional RAN by establishing an open framework that enables data-driven optimization at the individual user level, which is essential for the evolution of the next-generation (NextG) cellular networks. Accurate predictions of CPU demand for each user's equipment (UE) in the O-Cloud at the next step will enable more efficient CPU resource optimization. While AI is promising to optimize CPU utilization, it encounters two significant challenges. First, the varying number of UEs leads to shifts in feature dimensions, rendering the model unable to accept these inputs since the input dimension of the AI model remains fixed. Second, the ongoing introduction of new types of UEs over time with distinct CPU demands and dynamic combinations of various active devices adds further variability, thereby complicating predictive accuracy. To address the first challenge, in this research, we propose a novel dynamic number of devices management (DNDM) framework that effectively accommodates a dynamic number of devices in O-RAN, addressing the challenges associated with variable UE demands in future NextG O-RAN. We formulate an optimization problem for the second challenge, enabling the model to learn new demand scenarios while preserving knowledge from previously encountered configurations. To solve the optimization, we propose an exemplar replaybased continual adaptation (CA) framework designed to operate within the near real-time RAN Intelligence Controller (RT-RIC). The CA-DNDM actively prevents catastrophic forgetting and delivers continuous adaptability, seamlessly handling evolving UE types and quantities. Through extensive experimental results, we demonstrate that the proposed CA-DNDM framework effectively handles scenarios with varying UE counts and reliably predicts CPU demand for new situations while preserving the knowledge gained from prior scenarios.

Deep Joint Source-Channel Coding (Deep-JSCC) has emerged as a promising semantic communication approach for wireless image transmission by jointly optimizing source and channel coding using deep learning techniques. However, traditional Deep-JSCC architectures employ fixed encoder-decoder structures, limiting their adaptability to varying device capabilities, real-time performance optimization, power constraints and channel conditions. To address these limitations, we propose DD-JSCC: Dynamic Deep Joint Source-Channel Coding for Semantic Communications, a novel encoder-decoder architecture designed for semantic communication systems. Unlike traditional Deep-JSCC models, DD-JSCC is flexible for dynamically adjusting its layer structures in real-time based on transmitter and receiver capabilities, power constraints, compression ratios, and current channel conditions. This adaptability is achieved through a hierarchical layer activation mechanism combined with implicit regularization via sequential randomized training, effectively reducing combinatorial complexity, preventing overfitting, and ensuring consistent feature representations across varying configurations. Simulation results demonstrate that DDJSCC enhances the performance of image reconstruction in semantic communications, achieving up to 2 dB improvement in Peak Signal-to-Noise Ratio (PSNR) over fixed Deep-JSCC architectures, while reducing training costs by over 40%. The proposed unified framework eliminates the need for multiple specialized models, significantly reducing training complexity and deployment overhead.

Federated learning (FL) enables privacy-preserving model training across decentralized data. However, significant data heterogeneity, common in domains like the Internet of Things (IoT), hinders generalization. Federated Domain Generalization (FDG) extends the FL paradigm by aiming to train models that generalize effectively to unseen domains, without requiring access to data from those domains during training. Current FDG methods primarily use ResNet backbones pre-trained on ImageNet-1K, limiting adaptability due to architectural constraints and limited pre-training diversity. This reliance has created a gap in leveraging advanced architectures and diverse pre-training datasets to address these challenges. To bridge this gap, we present the first comprehensive investigation into the efficacy of advanced pre-trained architectures such as Vision Transformers, ConvNeXt, and Swin Transformers, in enhancing FDG performance. Unlike ResNet, these architectures capture global context and long-range dependencies, making them well-suited for FDG. Beyond architectural evaluation, we systematically assess the impact of diverse pre-training datasets and compare self-supervised and supervised strategies. Our analysis rigorously investigates the influence of architectural depth, parameter efficiency, and the interplay between diverse model families and dataset characteristics on FDG performance. We find that advanced architectures pre-trained on large datasets significantly outperform ResNet models. Specifically, ConvNeXt architectures outperform all other candidates. We find self-supervised methods using masked image patch reconstruction via discrete token prediction outperform their supervised counterparts. We observe that certain advanced model variants with fewer parameters outperform larger ResNet models. This underscores the need for advanced architectures and scalable pretraining to enable efficient and generalizable FDG.

The future cellular networks, known as Next Generation (NextG), are anticipated to employ infrastructure based on cloud computing, with programmable, virtualized, and disaggregated designs. The separation of control functions from physical infrastructure will be implemented, and the utilization of standardized interfaces will facilitate the establishment of tailored closed-control loops. The O-RAN (Open Radio Access Network) paradigm aims to resolve the issue of limited control over the RAN by proposing an open design framework that enables data-driven and intelligent optimization at the individual user level. The growing ubiquity of O-RAN networks has underscored the significance of network traffic categorization in effectively managing O-RAN networks and preserving cybersecurity. This article analyzes the O-RAN alliance’s disaggregated network architecture, focusing on its significant contribution to NextG networks. This paper proposes a novel approach to examine and assess traffic patterns inside the O-RAN architecture. The ongoing acquisition of knowledge regarding encrypted traffic is of utmost importance in light of the perpetual advancements in applications and the advent of encryption technologies. The ability to dynamically adjust in traffic analysis is crucial for effectively addressing the evolving network environment, specifically inside the O-RAN architecture. To address this, we propose a new method named Incremental O-RAN Traffic Categorization (IORTC), specially designed to categorize encrypted data within the O-RAN framework. The IORTC system manages the continuous collection of knowledge from encrypted traffic in the O-RAN framework. It is designed to adapt to the evolution of various encrypted traffic categories while retaining information about earlier encrypted traffic. Based on the experimental results, our method demonstrates a remarkable ability to assimilate new traffic patterns while managing the retention of previously acquired knowledge, and it performed well in all evaluation criteria with an average accuracy of 98%.

In the rapidly advancing landscape of 6G wireless communication, beamforming plays a crucial role especially with the utilization of millimeter-wave and terahertz frequency bands being pivotal for achieving ultra-high data rates. Despite their promise, these bands present a significant challenge due to the beam training overheads required for precise beamforming, particularly in high-mobility applications like intelligent transportation systems and emerging virtual reality platforms such as the metaverse. While initial deep learning models mitigates the beam training overheads, their performance is compromised due to sensitivity to environmental and lighting conditions, revealing a critical gap in robustness. To address this limitation, this paper proposed a semantic-based method specifically designed to enhance the robustness of beamforming. Utilizing the cutting-edge You Only Look Once version 8 (YOLOv8) algorithm, semantic data from RGB camera images has been extracted to significantly improve the system’s adaptability across a range of environmental conditions. Further, to complement the beam management a novel approach is proposed for the identification of target vehicle by employing K-means clustering in conjunction with the GPS data of the target vehicle. For maintaining the ultra reliable low latency communication (URLLC) a lightweight model has been used to predict the optimal beamforming index. The efficacy of our proposed model is empirically substantiated through rigorous experimental trials in real-world 6G environment, demonstrating significant improvements in average received power ranging from 6.49% to 38.27%, compared to the baselines.