Analyzing the Impact of COVID-19 Control Policies on Campus Occupancy and Mobility via WiFi Sensing

Camellia Zakaria, Amee Trivedi, Emmanuel Cecchet, Michael Chee, Prashant J. Shenoy, Rajesh Balan

TL;DR

This paper demonstrates how readily available WiFi infrastructure can be leveraged to monitor campus occupancy and mobility changes in response to COVID-19 policies. By analyzing anonymized WiFi connection logs, the authors show that coarse-grained data can effectively reveal the impact of measures like online learning and staggered schedules on crowd density and movement, offering a privacy-preserving approach for institutional safety management.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for effective public health measures, including contact tracing and social distancing. While many digital solutions emerged, relying on Bluetooth or GPS, these often face adoption challenges or limitations in indoor environments. Institutions, particularly universities, needed a way to monitor crowd behavior and ensure compliance with evolving safety guidelines without invasive surveillance. This paper addresses the challenge of using existing, ubiquitous WiFi infrastructure as a practical and privacy-conscious method to understand how COVID-19 control policies affected campus occupancy and mobility.

Methods

The core of the methodology involves analyzing WiFi network logs generated when smartphones connect to access points (APs). This "network-centric sensing" approach leverages existing infrastructure, requiring no special apps or user permissions.

  • Data Collection: WiFi logs were collected from three universities (two in Singapore, one in the US). These logs included timestamps, AP identifiers, and anonymized (hashed) MAC addresses of connected devices.
  • Occupancy Inference: Occupancy was determined by counting the unique number of associated WiFi devices within a specific area (building, floor, or room) at a given time. Unassociated devices (those scanning but not connected) were also considered for visitor analysis.
  • Mobility Inference: Mobility was tracked by observing the sequence of AP connections made by a device over time. This allowed for the inference of transitions between different areas, floors, and buildings.
  • Privacy Safeguards: All MAC addresses were anonymized using SHA-2 hashing. The analysis focused on aggregated data, ensuring no individual users were identified.

Datasets

The study utilized datasets collected from the production WiFi networks of three universities:

  • Singapore Management University (SMU)
  • National University of Singapore (NUS)
  • University of Massachusetts Amherst (UMASS)

The data spanned periods before, during, and after the implementation of various COVID-19 control policies, allowing for comparative analysis.

Key Insights and Ideas

  • WiFi as a Practical Sensor: The paper strongly advocates for WiFi sensing as a feasible and scalable solution for institutional monitoring, overcoming adoption barriers of other technologies.
  • Coarse-Grained Data Sufficiency: Surprisingly, relatively coarse WiFi data (AP-level location) is sufficient to capture significant changes in occupancy and mobility patterns, reflecting the impact of broad policy changes.
  • Policy Impact on Occupancy: Online learning and staggered schedules demonstrably reduced overall campus occupancy, with significant drops observed across all universities.
  • Policy Impact on Mobility: While occupancy dropped, individual mobility patterns showed varied responses. Strict lockdowns were most effective in reducing overall movement, but even then, some individuals continued to move within their immediate vicinity.
  • Dormitory Dynamics: In residential campuses, lockdown measures led to increased occupancy within dormitories, shifting activity from academic areas to living spaces.

Key Results

  • Occupancy Reduction: SMU experienced a >90% drop in occupancy from pre-COVID-19 to Phase 3/4 policies. NUS also saw significant reductions, though slightly less pronounced due to more students remaining in dorms. UMASS also showed a >90% decrease in occupancy between "business as usual" and "stay-at-home" orders.
  • Mobility Changes: While overall transitions decreased with stricter policies, the mobility rate per person remained relatively high in some cases until full lockdowns were enacted. This suggests that while people moved less overall, those who did move often visited similar locations.
  • Area-Specific Occupancy: Analysis revealed that recreational and dining areas saw significant drops in occupancy, while libraries and dormitories became more prominent locations during lockdown phases.

Limitations and Future Work

  • Under-counting: The study relies on WiFi-enabled devices that are associated with the network, potentially under-counting individuals without such devices or who do not connect.
  • Spatial Granularity: WiFi sensing provides coarse-grained location data, making it difficult to determine precise inter-user distances or identify individuals at immediate risk of exposure in small spaces.
  • Geographic Specificity: The findings are based on data from two countries with distinct policy approaches. While general trends were observed, local factors might influence applicability to other regions.

Future work could involve:

  • Deeper and broader analyses to address these shortfalls.
  • Exploring more sophisticated methods for inferring occupancy and mobility.
  • Investigating the impact of policies in other types of institutions or regions.
  • Developing more granular risk assessment models based on WiFi data.

Summary updated

BibTeX
@article{zakaria2022analyzing,
  author = {Camellia Zakaria and Amee Trivedi and Emmanuel Cecchet and Michael Chee and Prashant J. Shenoy and Rajesh Balan},
  title = {{Analyzing the Impact of COVID-19 Control Policies on Campus Occupancy and Mobility via WiFi Sensing}},
  journal = {ACM Trans. Spatial Algorithms Syst.},
  year = {2022},
  doi = {10.1145/3516524},
}
Figure