Jianchun Liu

Federated fine-tuning (FedFT) enables adapting pre-trained large vision models (LVMs) on distributed, privacy-sensitive devices, while its practical deployment is hindered by three critical challenges: resource constraints, system heterogeneity, and non-IID data. While prior studies partially address these issues, e.g., by pre-training initial models on synthetic images to mitigate the adverse effects of non-IID data, or leveraging parameter-efficient fine-tuning (PEFT) methods like low-rank adaptation (LoRA) to reduce resource consumption, they remain inadequate and fragmented. Specifically, existing synthetic image generation methods fail to capture device-specific feature distributions, while current PEFT-based FedFT methods often undervalue weaker devices that may provide critical information. More importantly, the separate optimization of pre-training and FedFT neglects their inherent connection, lacking a holistic perspective to maximize training efficiency. To overcome these limitations, we propose FeDiSyn, a unified framework that holistically considers the interplay between pre-training and FedFT to minimize the overall LVM training time. Specifically, FeDiSyn introduces: (i) a scaling law for FedFT pre-training to determine the optimal number of synthetic images, balancing pre-training benefit against generation/pre-training cost, (ii) diffusion-based synthetic image generation that captures device-specific feature distributions for pre-training to tackle non-IID data, and (iii) a contribution-aware LoRA configuration and bandwidth allocation algorithm for FedFT to ensure that informative devices are effectively utilized while addressing system heterogeneity. Experimental results on the real-world testbed demonstrate that FeDiSyn reduces completion time by over 52.5% and communication cost by over 97.2%, while achieving comparable accuracy to state-of-the-art solutions.

Compression is an efficient way to relieve the tremendous communication overhead of federated learning (FL) systems. However, for the existing works, the information loss under compression will lead to unexpected model/gradient deviation for the FL training, significantly degrading the training performance, especially under the challenges of data heterogeneity and model obsolescence. To strike a delicate trade-off between model accuracy and traffic cost, we propose Caesar, a novel FL framework with a low-deviation compression approach. For the global model download, we design a greedy method to optimize the compression ratio for each device based on the staleness of the local model, ensuring a precise initial model for local training. Regarding the local gradient upload, we utilize the device's local data properties (i.e., sample volume and label distribution) to quantify its local gradient's importance, which then guides the determination of the gradient compression ratio. We have implemented Caesar, on two physical platforms with 40 smartphones and 80 NVIDIA Jetson devices. Extensive results show that Caesar, can reduce the traffic costs by about 25.54%þicksim37.88% when achieving the same target accuracy compared to the compression-based baselines, while incurring only a 0.68% degradation in final test accuracy relative to the full-precision communication.

Federated fine-tuning (FedFT) of large language models (LLMs) has emerged as a promising solution for adapting models to distributed data environments while ensuring data privacy. Existing FedFT methods predominantly utilize parameter-efficient fine-tuning (PEFT) techniques to reduce communication and computation overhead. However, they often fail to adequately address the catastrophic forgetting, a critical challenge arising from continual adaptation in distributed environments. The traditional centralized fine-tuning methods, which are not designed for the heterogeneous and privacy-constrained nature of federated environments, struggle to mitigate this issue effectively. Moreover, the challenge is further exacerbated by significant variation in data distributions and device capabilities across clients, which leads to intensified forgetting and degraded model generalization. To tackle these issues, we propose FedBE, a novel FedFT framework that integrates an adaptive transformer block expansion mechanism with dynamic trainable-block allocation. Specifically, FedBE expands trainable blocks within the model architecture, structurally separating newly learned task-specific knowledge from the original pre-trained representations. Additionally, FedBE dynamically assigns these trainable blocks to clients based on their data distributions and computational capabilities. This enables FedBE to better accommodate heterogeneous federated environments and enhances the generalization ability of the model. Extensive experiments show that compared with existing federated fine-tuning methods, FedBE achieves 12-74% higher accuracy retention on general tasks after fine-tuning and a model convergence acceleration ratio of 1.9-3.1× without degrading the accuracy of downstream tasks.

WiFi-based gesture recognition holds great promise due to its non-intrusive and ubiquitous nature. However, robust generalization across diverse conditions remains a core challenge due to the inherent sensitivity and ambiguity of Channel State Information (CSI) signals, making them unreliable for real-world human-computer interaction (HCI) applications. While domain adversarial learning-based systems have proven effective in other sensing modalities, its performance in WiFi-based gesture recognition remains limited. In this paper, we revisit the foundational assumptions in current adversarial learning-based WiFi gesture recognition systems and identify a key mismatch: conventional methods rely on subjective physical domain labels (e.g., location, orientation), which often fail to reflect true distributional patterns in the data. This misalignment introduces two critical issues, namely classification conflict and manifold distortion, which fundamentally limit generalization performance. To address these challenges, we propose G es F i , a novel WiFi-based gesture recognition system that introduces WiFi latent domain mining to redefine domains directly from the data itself. G es F i first processes raw sensing data collected from WiFi receivers using CSI-ratio denoising, Short-Time Fast Fourier Transform, and visualization techniques to generate standardized input representations. It then employs class-wise adversarial learning to suppress gesture semantic and leverages unsupervised clustering to automatically uncover latent domain factors responsible for distributional shifts. These latent domains are then aligned through adversarial learning to support robust cross-domain generalization. Finally, the system is applied to the target environment for robust gesture inference. We deployed GesFi under both single-pair and multi-pair settings using commodity WiFi transceivers, and evaluated it across multiple public datasets and real-world environments. Compared to state-of-the-art baselines, G es F i achieves up to 78% and 50% performance improvements over existing adversarial methods, and consistently outperforms prior generalization approaches across most cross-domain tasks. (Code available at https://github.com/CamLoPA/GesFiCode).

With the increasing use of electronic devices, online handwriting verification has become crucial for biometricsbased identity authentication. Traditional methods, which rely on content-dependent verification of the writer's name, are vulnerable to forgery. This paper introduces a content-independent handwriting authentication system, Ph-Wri, designed for commodity smartphones. The core innovation is a multi-path attention feature fusion network that combines both static features (image of the handwritten text) and dynamic features (time-dependent properties during writing), to abstract the handwriting style instead of specific content for recognition, enabling robust user authentication. To extract handwriting style from dynamic writing features, we propose a polarity-aware attention strategy during training. This strategy incorporates Style Channel Attention (SCA) to capture direction-sensitive stylistic features, and Trajectory Spatial Attention (TSA) to highlight key handwriting trajectory regions. In the fine-tuning stage, the Correlation-Aware Attention (CAA) module models inter-channel structural correlations, mitigating the influence of content and enhancing style-consistent representations. By linking content-independent handwriting style to user identity, the system achieves accurate authentication. Extensive experiments on both the self-built CIEHD dataset and the public BiosecurID dataset demonstrate exceptional performance, achieving a 99% Verification Accuracy on CIEHD. Compared to state-of-theart methods that utilize only static or dynamic data, Ph-Wri significantly reduces the Equal Error Rate, showcasing the effectiveness and practicality of the proposed approach.

In this paper, we propose GesFi, a novel WiFi-based gesture recognition system that introduces WiFi latent domain mining to redefine domains directly from the data itself. GesFi first processes raw sensing data collected from WiFi receivers using CSI-ratio denoising, Short-Time Fast Fourier Transform, and visualization techniques to generate standardized input representations. It then employs class-wise adversarial learning to suppress gesture semantic and leverages unsupervised clustering to automatically uncover latent domain factors responsible for distributional shifts. These latent domains are then aligned through adversarial learning to support robust cross-domain generalization. Finally, the system is applied to the target environment for robust gesture inference. We deployed GesFi under both single-pair and multi-pair settings using commodity WiFi transceivers, and evaluated it across multiple public datasets and real-world environments. Compared to state-of-the-art baselines, GesFi achieves up to 78% and 50% performance improvements over existing adversarial methods, and consistently outperforms prior generalization approaches across most cross-domain tasks.

Federated learning (FL) enables training AI models across distributed edge devices (i.e., workers) using local data, while facing challenges including communication resource constraints, edge heterogeneity, and non-IID data. Over-the-air computation (AirComp) has emerged as a promising technique to improve communication efficiency by leveraging the superposition property of a wireless multiple access channel (MAC) for model aggregation. However, over-the-air aggregation requires strict synchronization among edge devices, which is essentially incompatible with the asynchronous FL mechanisms often used to handle edge heterogeneity. To overcome this incompatibility, we propose Air-FedGA, a grouping-based asynchronous FL mechanism via AirComp, where workers are organized into groups for synchronized over-the-air aggregation within each group, while groups asynchronously communicate with the parameter server to update the global model. This design retains the communication efficiency of AirComp while addressing training inefficiency caused by edge heterogeneity. We provide a rigorous convergence analysis for Air-FedGA, theoretically quantifying how the convergence bound depends on several key factors, such as the maximum staleness, the degree of non-IID data among groups, and the AirComp aggregation mean squared error (MSE). Guided by these theoretical insights, we propose power control and worker grouping algorithms to minimize the convergence bound by jointly optimizing the AirComp aggregation MSE and the grouping strategy. We conduct experiments on classical models and datasets, and the results demonstrate that our proposed mechanism and algorithms can accelerate the model training by 1.83- $2.22\times $ compared with the state-of-the-art solutions.