WiFi CSI
A Multi-Radio Audit of Subject-Dependent Performance in Recent Public WiFi CSI Datasets
Lead Researcher & Co-authorOverview
The work began as an attempt to use ESP32-based WiFi CSI sensing to estimate crowd density in rooms, with the longer-term idea of deploying such sensing across a college environment. During the investigation, we became interested in a more fundamental evaluation issue: WiFi CSI activity-recognition models can report unusually strong performance when random train/test splitting allows data from the same individuals or overlapping recording conditions to appear in both sets. This motivated a shift from collecting a new CSI dataset to auditing existing public WiFi CSI datasets under subject-aware evaluation.
Research Question
How much do reported WiFi CSI activity-recognition results depend on the choice of evaluation protocol - specifically, on whether train/test splits are random-window or subject-disjoint - and what does that say about subject and session leakage in recent public datasets?
My Contribution
- Initiated and drove the research direction from the original ESP32 CSI sensing idea toward an evaluation audit of existing public WiFi CSI datasets.
- Identified and sourced the datasets used in the study from IEEE DataPort.
- Helped formulate the central research problem around random-window versus subject-disjoint evaluation and potential subject/session leakage.
- Co-developed the preprocessing, evaluation, and analysis code with Sanjana Paudel.
- Participated directly in running experiments and interpreting the resulting performance differences across evaluation protocols.
- Took primary responsibility for organizing and framing the research into the final paper.
- Led manuscript preparation and the final technical review, working collaboratively with Sanjana.
Methodology
- Selected a set of recent public WiFi CSI activity-recognition datasets from IEEE DataPort that are representative of how the field reports results.
- Reproduced the standard random-window train/test splitting used in each dataset's original pipeline.
- Implemented a subject-disjoint evaluation protocol that keeps all samples from any given subject or recording session in a single split.
- Compared model performance across the two protocols and quantified the gap attributable to subject/session leakage.
Key Findings
- Random-window evaluation consistently produced substantially higher reported accuracy than subject-disjoint evaluation.
- The performance gap indicates that part of the strong results in these datasets is driven by subject or session leakage rather than true generalization.
- Subject-aware evaluation provides a noticeably more honest estimate of how these models would behave on unseen people and conditions.
What I Learned
- How to design reproducible ML evaluation that does not overstate generalization.
- How dataset leakage and distribution shifts skew reported results.
- How to design rigorous experiments and apply research methodology end-to-end.
Limitations
- The audit is restricted to the public datasets that were available and selected, which limits the breadth of the conclusions.
- Shared-subject leakage cannot always be fully distinguished from session-level overlap in every dataset.