Scale What Counts, Mask What Matters: Evaluating Foundation Models for Zero-Shot Cross-Domain Wi-Fi Sensing

TL;DR

This work tackles the persistent domain-shift problem in Wi-Fi sensing by adopting a foundation-model paradigm trained with masked autoencoding on a large, diverse corpus of CSI data. By pretraining a ViT-based encoder on 1.3M samples from 14 heterogeneous datasets and evaluating across cross-domain tasks, the study reveals data scale and diversity as the primary drivers of cross-domain performance, with model capacity offering diminishing returns under current data constraints. It demonstrates that large-scale pretraining improves cross-domain accuracy by about 2.2% to 15.7% across human activity, gesture, and user-identification tasks, and that diversity enables robust zero-shot transfer even when target-domain data are excluded from pretraining. The findings provide actionable guidance for designing Wi-Fi sensing systems that generalize to real-world deployments, emphasizing data aggregation, expressive input representations, and efficient training strategies to balance performance with computational costs.

Abstract

While Wi-Fi sensing offers a compelling, privacy-preserving alternative to cameras, its practical utility has been fundamentally undermined by a lack of robustness across domains. Models trained in one setup fail to generalize to new environments, hardware, or users, a critical "domain shift" problem exacerbated by modest, fragmented public datasets. We shift from this limited paradigm and apply a foundation model approach, leveraging Masked Autoencoding (MAE) style pretraining on the largest and most heterogeneous Wi-Fi CSI datasets collection assembled to date. Our study pretrains and evaluates models on over 1.3 million samples extracted from 14 datasets, collected using 4 distinct devices across the 2.4/5/6 GHz bands and bandwidths from 20 to 160 MHz. Our large-scale evaluation is the first to systematically disentangle the impacts of data diversity versus model capacity on cross-domain performance. The results establish scaling trends on Wi-Fi CSI sensing. First, our experiments show log-linear improvements in unseen domain performance as the amount of pretraining data increases, suggesting that data scale and diversity are key to domain generalization. Second, based on the current data volume, larger model can only provide marginal gains for cross-domain performance, indicating that data, rather than model capacity, is the current bottleneck for Wi-Fi sensing generalization. Finally, we conduct a series of cross-domain evaluations on human activity recognition, human gesture recognition and user identification tasks. The results show that the large-scale pretraining improves cross-domain accuracy ranging from 2.2% to 15.7%, compared to the supervised learning baseline. Overall, our findings provide insightful direction for designing future Wi-Fi sensing systems that can eventually be robust enough for real-world deployment.

Figures (7)

• Figure 1: Foundation Model powering various Wi-Fi sensing tasks (Image generated by Google Gemini).
• Figure 2: Overview of our Wi-Fi sensing foundation model evaluation. We collect a large-scale and diverse set of Wi-Fi CSI from public datasets that cover various tasks, devices, frequency bands, and environments. After harmonizing and preprocessing the data, we pretrain a Masked Autoencoder (MAE) model on the unified CSI representation. The pretrained encoder can then be fine-tuned on downstream sensing tasks with limited labels.
• Figure 3: illustration of typical artifacts in CSI data
• Figure 4: Visualization of original and reconstructed CSI samples. Despite heavy masking, the model effectively reconstructs the missing patches, capturing both temporal dynamics and frequency patterns essential for Wi-Fi sensing tasks.
• Figure 5: Scaling behaviors of Wi-Fi sensing foundation models under cross-domain evaluation on Widar 3.0 dataset.