Multimodal beam prediction leverages roadside sensors to reduce beam training overhead in millimeter-wave (mmWave) Vehicle-to-Infrastructure (V2I) communications. However, representative missing-modality methods primarily address offline learning with incomplete multimodal inputs, rather than continual adaptation to post-deployment distribution shifts. Moreover, these methods suffer from catastrophic forgetting when fine-tuned to adapt to new data distributions, limiting their applicability to practical V2I deployments. In this paper, we propose a multimodal beam prediction framework that combines Mixture-of-Experts (MoE) fusion with dual Low-Rank Adaptation (LoRA), achieving robustness under missing modalities and distribution shift. Our method adopts a two-stage training strategy. Specifically, we design a Flexible Embedding Bank that maintains distinct learnable embeddings indexed by the combination of available modalities, providing combination-specific substitutes for MoE fusion under missing modalities. Furthermore, we employ dual LoRA, which separates low-rank adapters into a frozen branch that accumulates consolidated knowledge and a trainable branch that learns new patterns, enabling adaptation without forgetting through additive merging. Experimental results on the DeepSense 6G dataset demonstrate competitive performance against existing methods. Specifically, our framework improves performance by 17% under missing modalities and achieves 96% top-5 accuracy on the target adaptation set with only 0.34% of the trainable parameters required for full fine-tuning.
Cell-free networks enhance wireless coverage and capacity by allowing user equipments (UEs) to be served by multiple access points (APs). To ensure scalability, user-centric cell-free networks reduce overhead by selecting a subset of APs for each UE. However, existing AP selection methods rely on a vast amount of measurements, leading to high feedback overhead and processing delays. To address this challenge, we propose a meta spatial graph neural network (S-GNN) model to improve AP selection accuracy with limited measurements and training data, by training on multiple AP deployment scenarios and fine-tuning on a single AP deployment scenario. In order to handle temporal variations in high-mobility scenarios, we introduce a meta spatial-temporal GNN (ST-GNN) model that integrates gated recurrent units (GRUs) to predict future AP selections. Simulation results demonstrate that the proposed framework consistently outperforms its state-of-the-art benchmarks, reducing feedback overhead considerably while maintaining high accuracy, which positions our solution as a competitive candidate for AP selection in user-centric cell-free networks.
The increasing demand for high-resolution visual services introduces significant pressure on bandwidth-limited edge-device networks, making collaborative super-resolution (SR) an alternative to high-resolution content transmission. However, edge-device collaborative SR offloading is challenging due to heterogeneous computational capabilities and uncertainty in both network conditions and offloading decisions, leading to degraded efficiency. To address these challenges, we propose CoCA-SR, a conflict-aware collaborative SR framework for edge-device networks. CoCA-SR operates at the tile level to support fine-grained computation offloading, where lightweight SR models are assigned to local devices while computation-intensive regions are offloaded to the edge server. Specifically, we consider two types of conflicts arising from tile to edge offloading and tile to device offloading. For edge-side offloading, bandwidth uncertainty cause inaccurate resource availability estimation and lead to conflicting offloading decisions. We therefore integrate bandwidth estimation into a Double Deep Q-Network (Double-DQN) to enhance edge-side scheduling. Furthermore, we design a conflict-aware multi-armed bandit (CAMAB) framework, where each requesting device independently learns the effective tile-serving capability of serving devices from aggregate reward feedback under different numbers of assigned tiles. This enables decentralized offloading without any communication among requesting devices, while implicitly capturing conflict through reward feedback. Extensive experiments under diverse network conditions demonstrate that CoCA-SR outperforms state-of-the-art (SOTA) edge-assisted and collaborative SR schemes in reconstruction quality and offloading efficiency.
Semantic communication has emerged as a transformative paradigm for improving communication efficiency, but its security has been largely overlooked, rendering semantic codec models highly vulnerable to adversarial example attacks. To address this, we propose a detection–purification–restoration defense (DPRD) framework for image semantic communication systems. A detector based on feature distribution separation first identifies adversarial examples to prevent their direct transmission; the residual denoising diffusion model (RDDM) then purifies them while recovering semantic information; and an image detail restoration module (IDRM) finally removes residual perturbations and restores fine-grained image details. Experimental results under four mainstream attack methods demonstrate that DPRD outperforms six baseline defense methods, generalizes well to unseen attack types, different channels, and different communication systems, while enabling plug-and-play deployment with low overhead on clean inputs. This work provides a comprehensive evaluation of adversarial defense in semantic communication and showcases its potential for enhancing the security and reliability of next-generation intelligent communication networks.
Direct-to-Cell (D2C) Satellite-Terrestrial Integrated Networks (STINs) are emerging as a promising technology to provide ubiquitous global connectivity. To support direct access for existing smartphones, D2C satellite systems must reuse spectrum bands originally allocated to terrestrial cellular networks, which may introduce severe co-frequency interference. To address this issue, this paper investigates the interference-aware resource allocation problem in D2C-enabled STINs, with the objective of maximizing the system’s overall Energy Efficiency (EE). To efficiently solve the resulting non-convex optimization problem, we propose a multi-stage framework which decomposes the original problem into three subproblems: user association, power allocation and subcarrier allocation. To reduce time complexity, a spatial clustering strategy is employed to significantly reduce the number of decision variables while determining user association. After that, we utilize Alternating Optimization (AO) to iteratively solve the power and subcarrier allocation subproblems. For power allocation, we introduce an adaptive rate decomposition technique to decouple the minimum rate constraints. Furthermore, we apply a linear approximation to simplify the interference terms within the framework of Dinkelbach’s method and Successive Convex Approximation (SCA), enabling the derivation of closed-form solutions. For the subcarrier allocation subproblem, we derive an efficient greedy rule tailored for the decoupled separable per-subcarrier assignment. Extensive numerical simulations demonstrate that the proposed algorithm outperforms benchmark schemes across diverse scenarios.
