Qimei Chen

With the integration of sensing, communication, and computation (ISCC) emerging as a key paradigm for 6G, feder ated learning (FL) with ISCC (FL-ISCC) provides a promising solution to enhance training efficiency by orchestrating sample collection, local model updating, and wireless exchange. This work focuses on two core FL-ISCC algorithms: Federated Av eraging (FedAVG)-ISCC and Federated Stochastic Gradient De scent (FedSGD)-ISCC. Although both are widely used, their com parative performance and theoretical characterization remain inadequately explored. To address this, we establish a general FL ISCC framework and conduct a comprehensive theoretical and experimental analysis. Our main contributions include deriving the convergence rates for both algorithms under realistic network conditions, accounting for data heterogeneity and communication errors. The analysis reveals that FedAVG-ISCC achieves higher accuracy under independent and identically distributed (IID) data, while FedSGD-ISCC exhibits stronger robustness in non IID and error-prone wireless environments. These theoretical findings are corroborated by numerical simulations, providing actionable insights for selecting and deploying FL algorithms in practical ISCC systems.

Federated Edge Learning (FEL) enables privacy-preserving collaborative model training across distributed edge devices, yet its efficacy is critically challenged by the dual constraints of Non-IID data distributions and limited on-device resources. While existing methods often address communication, computation, or data heterogeneity in isolation, a holistic solution remains underexplored. To bridge this gap, we propose LightFEL, a novel FEL framework that incorporates adaptive model slimming via dynamic layer freezing. By selectively freezing convolutional layers during local training based on their sensitivity to data heterogeneity, LightFEL eliminates unnecessary backward propagation, thereby significantly reducing computational load and communication overhead, while inherently enhancing robustness to Non-IID data. We formulate a joint optimization problem that co-designs layer freezing policies with communication and computation resource allocation to minimize training latency. A theoretical convergence analysis is provided, establishing the interplay between sparsification strategies, resource decisions, and learning performance under Non-IID settings. The resulting mixed-integer nonlinear program is efficiently solved using a Q-learning and Lagrange Multiplier approach. Extensive experiments demonstrate that LightFEL substantially reduces computation costs and cuts communication overhead, while achieving higher accuracy and faster convergence under challenging Non-IID scenarios compared with benchmarks, which validate its dual capability in resource efficiency and heterogeneity robustness.

Fine-tuning (FT) large language models (LLMs) is crucial for adapting general-purpose models to specific tasks, enhancing accuracy and relevance with minimal resources. To further enhance generalization ability while reducing training costs, this paper proposes Federated LoRA with Dropout (FedLoDrop), a new framework that applies dropout to the rows and columns of the trainable matrix in Federated LoRA. A generalization error bound and convergence analysis under sparsity regularization are obtained, which elucidate the fundamental trade-off between underfitting and overfitting. The error bound reveals that a higher dropout rate increases model sparsity, thereby lowering the upper bound of pointwise hypothesis stability (PHS). While this reduces the gap between empirical and generalization errors, it also incurs a higher empirical error, which, together with the gap, determines the overall generalization error. On the other hand, though dropout reduces communication costs, deploying FedLoDrop at the network edge still faces challenges due to limited network resources. To address this issue, an optimization problem is formulated to minimize the upper bound of the generalization error, by jointly optimizing the dropout rate and resource allocation subject to the latency and per-device energy consumption constraints. To solve this problem, a branch-and-bound (B&B)-based method is proposed to obtain its globally optimal solution. Moreover, to reduce the high computational complexity of the B&B-based method, a penalized successive convex approximation (P-SCA)-based algorithm is proposed to efficiently obtain its high-quality suboptimal solution. Finally, numerical results demonstrate the effectiveness of the proposed approach in mitigating overfitting and improving the generalization capability.

Powered by Artificial Intelligence (AI), AI-Generated Content (AIGC) has recently emerged as a promising approach for synthesizing multimodal content, including text, images, and videos. Diffusion model is one of the predominant frameworks within AIGC for generating diverse and photorealistic images with high fidelity. However, deploying diffusion models on practical resource-constrained edge devices confronts critical challenges for their limited computational abilities and severe communication overhead, which spurs research interests in distributed AIGC. Existing works on distributed AIGC primarily focus on either increasing training cost to pursue higher generation quality or sacrificing performance to achieve communication efficiency, merely considering their integration impacts. To address these issues, we propose a novel Cross-Modal Collaborative Diffusion Model (Co-Diff) framework that achieves computation-and-communication efficient image synthesis without compromising robust generalization. The core innovation lies in our split diffusion architecture, which strategically offloads computation-intensive reverse denoising to the server and edge devices perform lightweight forward diffusion with text-guided attention. In addition, we design a learnable quantization module that encodes high-dimensional noise into compact codes to minimize communication overhead. We establish a theoretical convergence analysis for Co-Diff, deriving a closed-form expression that reveal a quantifiable trade-off between denoising accuracy and computational latency. Based on the theoretical findings, we formulate a joint communication-and-computation optimization problem, which obtains an analytical solution among diffusion-step scheduling, processing-frequency allocation, and bandwidth distribution. Extensive experiments validate our theoretical analysis, and demonstrate the effectiveness of the proposed design for computation-and-communication efficient Co-Diff. Compared with conventional distributed learning frameworks, the proposed Co-Diff increases communication-and-computation efficiency by 72.8%, as well as boosts image synthesis quality by 12.9%.

Wireless sensing has recently found widespread applications in diverse environments, including homes, offices, and public spaces, utilizing Channel State Information (CSI). However, CSI is highly sensitive to environmental changes, where even minor alterations can significantly distort the CSI patterns. This sensitivity often leads to performance degradation when applying models trained in one environment to another. To address this challenge, Domain Alignment Learning (DAL) has been widely adopted for cross-domain tasks, as it focuses on aligning the global distributions of the source and target domains in feature space. Despite its popularity, DAL often neglects inter-category relationships, which can lead to misalignment between categories across domains, even when global alignment is achieved. To overcome it, we propose KNN-MMD, a novel few-shot method for cross-domain wireless sensing. Our approach begins by constructing a "help set" using K-Nearest Neighbors (KNN) from the target domain, enabling local alignment between the source and target domains within each category using Maximum Mean Discrepancy (MMD). Additionally, most existing methods struggle to determine an optimal stopping point during training due to the absence of labeled data from the target domain. Our method resolves this by excluding the support set from the target domain during training and employing it as a validation set to determine the stopping criterion. We evaluate the proposed method across several cross-domain Wi-Fi sensing tasks, including gesture recognition, person identification, fall detection, and action recognition. In a one-shot scenario, our method achieves accuracy rates of 93.26%, 81.84%, 77.62%, and 75.30% for the respective tasks.

Network anomaly detection is widely used in network analysis and security prevention, in which reconstruction-based approaches have achieved remarkable results. However, attributed networks exhibit highly nonlinear relationships and time dependence over time, which make the anomalies more complex and ambiguous, resulting in anomaly detection still facing challenges. To this end, this study proposes an adversarial reconstruction framework with spectral-augmented and graph joint embedding for anomaly detection (GAN-SAGE), which integrates an autoencoder (AE) based on the frequency feature enhanced graph transformer (GT) into the generator for generating adversarial networks (GAN), improving network representation through adversarial training. The first stage of the encoding process captures the frequency domain information of the input timing data through spectral-augmented, and the second stage enhances the modeling capability of spatial structure and graph interaction dependency through multi-attribute coupling and GTs. We conducted extensive experiments on AIOps, SWaT and WADI datasets, demonstrating the effectiveness of GAN-SAGE compared to the state-of-the-art method. The detection performance of GAN-SAGE, respectively, improved by an average of 9.64%, 18.73% and 19.79% in terms of F1-score across the three datasets.

With the prevalence of emerging artificial intelligence services in next-generation wireless edge networks, Split Federated Learning (SFL), which divides a learning model into server-side and client-side models, has emerged as an appealing technology to deal with the heavy computational burden for network edge clients. However, existing SFL frameworks would frequently upload smashed data and download gradients between the server and each client, leading to severe communication overheads. To address this issue, this work proposes a novel communication-and-computation efficient SFL framework, which allows dynamic model splitting (server- and client-side model cutting point selection) and broadcasting of aggregated smashed data gradients. We theoretically analyze the impact of the cutting point selection on the convergence rate, revealing that model splitting with a smaller client-side model size leads to a better convergence performance and vise versa. Based on the above insights, we formulate an optimization problem to minimize the model convergence rate and latency under the consideration of data privacy via a joint Cutting point selection, Communication and Computation resource allocation (CCC) strategy. To deal with the proposed mixed integer nonlinear programming optimization problem, we develop an algorithm by integrating the Double Deep Q-learning Network (DDQN) with convex optimization methods. Extensive experiments validate our theoretical analyses across various datasets, and the numerical results demonstrate the effectiveness and superiority of the proposed communication-efficient SFL compared with existing schemes, including parallel split learning and traditional SFL mechanisms.

Federated multi-task learning (FMTL) is a promising technology to tackle one of the most severe non-independent and identically distributed (non-IID) data challenge in federated learning (FL), which treats each client as a single task and learns personalized models by exploiting task correlations. However, the transmission of individual task models generally results in a significant amount of communication overhead compared with global model broadcasting. Furthermore, related works mainly focus on FMTLs with default and static relationships among tasks, which obliterates the non-IID data characteristic. To address these issues, we propose a novel Clustered FMTL mechanism via Sparse Sharing (FedSS). Specifically, we introduce an iterative model pruning approach that trains customized client models to deal with the non-IID issue. Thereafter, we divide clients into different tasks according to their model similarities to promote communication efficiency. Based on clustered tasks, we introduce a sparse sharing mechanism that allows clients to share model parameters dynamically among different tasks to further boost the training performance. On the other aspect, the infertile communication resources would degrade the FMTL performance by restricting the personalized model transmissions. Hence, we first theoretically analyze the convergence performance of the proposed FedSS, which quantitatively unveils the relationship between the local model training performance and communication resources. Thereafter, we formulate a communication-and-computation efficient optimization problem via a joint sparsity ratio assignment and bandwidth allocation strategy. Closed-form expressions for the optimal sparsity ratio and bandwidth allocation are derived based on Lyapunov optimization and block coordinate update (BCU) algorithms. Numerical results illustrate that the proposed FedSS outperforms the benchmarks, and achieves an efficient communication and computation performance.

In recent years, Wi-Fi sensing has garnered significant attention due to its numerous benefits, such as privacy protection, low cost, and penetration ability. Extensive research has been conducted in this field, focusing on areas such as gesture recognition, people identification, and fall detection. However, many data-driven methods encounter challenges related to domain shift, where the model fails to perform well in environments different from the training data. One major factor contributing to this issue is the limited availability of Wi-Fi sensing datasets, which makes models learn excessive irrelevant information and over-fit to the training set. Unfortunately, collecting large-scale Wi-Fi sensing datasets across diverse scenarios is a challenging task. To address this problem, we propose CrossFi, a siamese network-based approach that excels in both in-domain scenario and cross-domain scenario, including few-shot, zero-shot scenarios, and even works in few-shot new-class scenario where testing set contains new categories. The core component of CrossFi is a sample-similarity calculation network called CSi-Net, which improves the structure of the siamese network by using an attention mechanism to capture similarity information, instead of simply calculating the distance or cosine similarity. Based on it, we develop an extra Weight-Net that can generate a template for each class, so that our CrossFi can work in different scenarios. Experimental results demonstrate that our CrossFi achieves state-of-the-art performance across various scenarios. In gesture recognition task, our CrossFi achieves an accuracy of 98.17% in in-domain scenario, 91.72% in one-shot cross-domain scenario, 64.81% in zero-shot cross-domain scenario, and 84.75% in one-shot new-class scenario. The code for our model is publicly available at https://github.com/RS2002/CrossFi.

The massive-antenna wideband millimeter wave (mmWave)/terahertz (THz) systems inevitably suffer from a severe beam split effect due to the non-negligible signal propagation delays, which dramatically reduces communication efficiency. Nevertheless, if the wideband split effect is properly utilized, it can also bring benefits via sensing split directions for channel training. Hence, this paper proposes a novel wideband beam alignment framework with true-time-delayer (TTD) modules, which can fully exploit the controllable split beams for efficient angle-of-arrivals (AoAs) estimation. Moreover, we develop a hierarchical posterior matching (PM) enabled wideband beam alignment approach, which proactively configures the split beams to accelerate the estimation of AoAs posterior probability distributions. To deal with the computational complexity of the predesigned codebook and the insensitivity of the Gaussian distribution assumption in PM, we further introduce a low-complex and high-flexible wideband beam alignment approach based on a deep unfolding mechanism. Numerical results verify that: 1) The proposed framework can significantly improve the AoAs estimation accuracy at the cost of the same pilot overheads. 2) The proposed low-complexity deep unfolding approach outperforms the conventional PM mechanism even in low signal-to-noise-ratio (SNR) scenarios.