Commodity WiFi sensing has various advantages, but it suffers from timing and frequency offsets due to non-synchronized transceiver and hardware imperfections. The main challenge is to estimate and compensate the residual carrier frequency offset (CFO), because the CFO and target motion are superimposed in the received channel state information (CSI) phase and cannot be effectively separated by conventional methods. In this paper, a deep learning-based CFO estimation method is proposed for suppressing the interference induced by millimeter-level vibrations, requiring only a single antenna. Mathematical derivations are first performed to analyze the influence of motion on the received CSI, revealing that the motion-induced interference in the CFO estimate can be eliminated by the scaled CSI amplitude. The scaling factor of CSI amplitude is determined adaptively by two well-designed convolutional neural networks (CNNs). This problem is modeled as a multi-class classification task with the objective of identifying the optimal scaling factor among all possible candidates. Experimental results verify that the proposed CFO estimation method exhibits robustness against millimeter-level vibrations. On a self-constructed dataset, the proposed method reduces the motion-induced interference amplitude by over 10 dB on average under the in-domain setting, and by over 6 dB under cross-domain settings, compared to the existing methods.
Reliable millimeter-wave (mmWave) vehicle-to-vehicle (V2V) communications depend on accurate beam tracking over highly directional links. Although multimodal sensing, including cameras, GPS, and radars, can provide valuable side information for beam prediction, practical deployments often suffer from missing modality conditions caused by sensor failures, blockage, or environmental occlusions, which severely degrade the performance. Existing solutions either rely on heuristic imputation that introduces severe distribution shifts, or adopt generic incomplete multimodal learning models that fail to adequately preserve the geometric discriminability required for beam tracking. To address these challenges, we propose CMGF-V2V, a novel Cross-Modal Generative Framework designed for robust V2V beam tracking under incomplete multimodal sensing. Specifically, CMGF-V2V adopts normalizing flows coupled with a direction-discriminative prior to align heterogeneous modalities into a common latent space. Subsequently, an attention-based fusion module is developed to infer the latent representations of missing modalities from available observations. Moreover, a conditional diffusion model is employed to capture modality-specific distributions and refine the generated features. Experiments on the real-world V2V datasets show that CMGF-V2V outperforms nine state-of-the-art incomplete multimodal learning baselines across distinct scenarios, demonstrating that the proposed framework achieves precise scene-consistent feature generation and superior robustness in beam tracking.
User-facing network applications generate diverse mix-flows with different performance requirements, such as low latency and high throughput. Classic heuristic congestion control algorithms (CCAs) struggle to perform well across diverse user preferences due to their fixed congestion control rules. Although learning-based CCAs can adapt to different user preferences, ensuring friendly coexistence among mix-flows with different user preferences is a key challenge. In this paper, we present Chameleon to ensure the friendly coexistence of learning-based CCA for such mix-flows. Our key insight is that CCA should adjust its preference based on the competitor type. Specifically, Chameleon employs a preference detection mechanism at the end host to identify the preferences of Chameleon-enabled competing flows. When a delay-preferred flow detects a throughput-preferred competing flow, it switches to a throughput-preferred policy to enhance its competitiveness and switches back to the delay-preferred policy after detecting the throughput-preferred flow leaves. Extensive evaluations in both large-scale emulation environments and real-world testbeds demonstrate that Chameleon enables friendly coexistence among mix-flows and improves the fairness index by up to 66% compared to existing CCAs.
Device-free fall detection using WiFi Channel State Information (CSI) offers a privacy-preserving approach to elderly health monitoring. Its performance can nevertheless decline in unseen environments because of static background overfitting and Non-Line-of-Sight (NLoS) signal attenuation. We present a domain-generalizable framework that combines a CSI-specific physics-driven front end with an attention-refined CNN-Transformer backbone. The Dynamic Variance Gate (DVG) uses local temporal variance to attenuate quiescent responses and retain motion-sensitive regions, while Physics-Aware Data Augmentation exposes the model to signal variations associated with deployment changes. EfficientNet-B0, CBAM, and the Transformer encoder are adapted to the CSI time-subcarrier representation: the backbone preserves a compact temporal feature map, CBAM recalibrates latent feature channels and temporal locations, and the Transformer models the ordered stages of a fall event. The aggregated-training NLoS subset analysis records 97.6% accuracy. Environment-disjoint evaluation in Home Lab C records 92.9% NLoS accuracy and 98.8% LoS accuracy without target-domain fine-tuning. Live field tests with 12 unseen participants and 200 events achieve 94.0% accuracy and 96.0% recall. The deployed pipeline evaluates an overlapping 5-second window every 0.5 seconds. On the field edge PC, a timing replay configured with approximately 7000 packets takes 0.08 s on average for preprocessing and all triggered inference calls, while each model inference averages 15 ms.
Multimodal foundation models (MFMs) support generalizable inference in distributed mobile sensing applications by integrating heterogeneous data streams. However, existing MFM inference pipelines typically require complete and synchronized multimodal inputs, whereas mobile sensing streams arrive asynchronously due to network variability and heterogeneous data volumes across modalities. Waiting for all modalities increases end-to-end latency, while inference with incomplete inputs can reduce accuracy. We argue that multimodal inference should depend on the utility of available information rather than modality completeness. Based on this principle, we propose AdaModalX, a progressive multimodal inference framework that decouples inference execution from modality synchronization through runtime utility-aware orchestration. At each device, importance-aware data uploading uses a lightweight probe distilled from the edge-side MFM to allow reliable samples to exit locally while uncertain samples continue uploading. At the edge, on-demand MFM activation constructs candidates from fully received samples and, when needed, imputed features derived from partially received delayed inputs. A lightweight contribution assessor built on frozen MFM encoders estimates candidate utility and activates the MFM once the candidate meets its corresponding utility threshold. Experiments on four mobile sensing applications show that AdaModalX improves the accuracy-latency trade-off, achieving up to 26.97% accuracy improvement and up to 64.75% end-to-end latency reduction.
