Network-side digital contact tracing on a large university campus

Suman Banerjee, Matthew L. Malloy, Lance Hartung, Steve Wangen, Suman Banerjee 0001MobiCom 2022 · DOI

TL;DR

This paper introduces a novel network-side digital contact tracing system that leverages WiFi connection logs to identify potential COVID-19 exposures on a large university campus. By analyzing AP colocation, the system demonstrates a predictive value of over 10% for identifying individuals who will test positive, offering a privacy-preserving alternative to app-based solutions.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for effective digital contact tracing to curb disease spread. Traditional methods often rely on mobile apps or operating system-level access, which face significant hurdles in terms of user adoption and privacy concerns. This paper addresses the challenge of implementing large-scale, automated contact tracing without requiring end-user installation or explicit consent, by utilizing existing network infrastructure. The authors aim to demonstrate the efficacy of using WiFi access point (AP) colocation data as a proxy for physical colocation and thus epidemiological contact.

Methods

The core of the proposed system is an AP colocation algorithm. This algorithm constructs a time-varying, weighted, undirected graph where nodes represent users (or their devices) and edges represent potential epidemiological contacts. The weight of an edge is a confidence score derived from the frequency and duration of simultaneous connections to the same WiFi AP.

A key innovation is the introduction of a cumulative exposure score. This score aggregates the confidence scores from all positive contacts an individual has had, aiming to capture the increased risk associated with exposure to multiple infected individuals. This score is then used to predict future positive cases.

The study also developed a method to infer a ground truth dataset for COVID-19 cases by analyzing WiFi association logs in designated isolation dormitories. This allowed for the validation of their AP-colocation and exposure score algorithms.

Datasets

The primary dataset used in this study consists of over 216 million WiFi client-access-point associations collected from more than 11,000 wireless access points (APs) across a large public university campus in the United States. This data was collected over the period of January 18, 2021, to May 7, 2021.

A ground truth dataset was inferred from this data, comprising 350 positive COVID-19 cases and 6,101 negative cases, based on their WiFi association patterns in isolation versus non-isolation residence halls.

Key Insights and Ideas

  • WiFi logs as a powerful proxy for contact tracing: The study demonstrates that readily available WiFi connection data can be effectively used to infer potential close contacts, bypassing the need for dedicated contact tracing apps.
  • AP colocation as a reliable indicator of proximity: The assumption that connecting to the same AP at the same time implies physical proximity and potential exposure is validated by the predictive power of the system.
  • Cumulative exposure score enhances prediction: Simply identifying direct contacts is insufficient. The cumulative exposure score, which accounts for exposure to multiple positive individuals, significantly improves the prediction of future positive cases.
  • Privacy-preserving by design: By anonymizing MAC addresses and relying on network-side data, the approach offers a more privacy-conscious alternative to app-based solutions.

Key Results

  • The AP-colocation approach has a predictive value of greater than 10%. Specifically, contacts identified by the algorithm have a greater than 10% chance of testing positive in the following 7 days, compared to a 0.79% chance for randomly selected individuals (a relative risk ratio of 12.6).
  • The cumulative exposure score can predict positive cases with a true positive rate of 16.5% and a missed detection rate of 79% at a specified operating point.
  • At a scale of 1.7 predicted contacts per positive case, the algorithm achieves a 10% positive predictive value (PPV). Increasing the scale to 5.2 contacts per case results in a 5.0% PPV.
  • The study successfully inferred ground truth COVID-19 cases and validated the system's performance against this inferred truth.

Limitations and Future Work

  • Device dependency: The system relies on individuals carrying a digital device that associates with the WiFi network. The percentage of individuals not using such devices is not accounted for.
  • False positives/negatives: AP colocation is a proxy for physical colocation, and there are inevitable false positives (individuals at the same AP but not in close proximity) and missed detections. The approach is best suited for identifying repeated and long-term interactions.
  • Privacy considerations: While the approach is designed to be privacy-preserving, the authors acknowledge that significant privacy risks must be addressed and balanced against the benefits.
  • Actionable interventions: Determining actionable and effective criteria for implementing interventions based on the system's predictions remains an outstanding challenge.
  • Future work: The authors suggest further work on determining if notification of possible exposure is warranted and exploring the extrapolation of these techniques to other settings where network log data is collected, such as cellular networks and digital advertising ecosystems. They also highlight the potential for finer-grained spatial resolution (e.g., room level) to improve accuracy.

Summary updated

BibTeX
@inproceedings{banerjee2022network,
  author = {Suman Banerjee and Matthew L. Malloy and Lance Hartung and Steve Wangen and Suman Banerjee},
  title = {{Network-side digital contact tracing on a large university campus}},
  booktitle = {MobiCom},
  year = {2022},
  doi = {10.1145/3495243.3517029},
}
Figure