Digital Contact Tracing (Part three): Technologies for Identifying Contact
Updated: Jul 16, 2020
In this blog series, my goal has been to explain contact tracing in an accessible manner. In part one, I discussed what contact tracing is and how technology can help enable it. In part two, I listed the metrics that one needs to evaluate any digital contact tracing solution. To jog your memories, the three metrics are: accuracy, privacy, and ubiquity. The technology should have low false-positive and false-negative rates (accuracy), it should afford users control over what data is shared with whom (privacy), and every user should be able to use this solution irrespective of what phone they use or what operating system they run (ubiquity). In this part, I will discuss some of the technological solutions that are being discussed to enable digital contact tracing and how I expect them to fare on the three metrics.
Some of the answers and evaluations will vary depending on the app design. However, some of these metrics vary fundamentally with the underlying technology being used to determine proximity. I will list the trade-offs associated with some common proximity detection technologies below. While there are several aspects of the solution like privacy-preserving communication, encryption, etc., I will focus on the question of proximity detection, i.e. how do you determine if two people are close to each other.
TL;DR: All the technologies have inherent trade-offs. See the table below.
Bluetooth has been a frustrating but useful feature on our smartphones for over a decade. Even in the pre-smartphone era, some feature phones used Bluetooth to transfer files. Bluetooth’s availability on a large range of phones has made it one of the most promising candidates for digital tracing. Recall how you need to put two devices close to each other to enable pairing. So, if two devices are close enough to form a connection, their owners may be close enough to transfer the COVID-19 virus. So, a bonus point to Bluetooth for ubiquity.
There are two levels of Bluetooth-based contact tracing. At the most basic level, just the presence of another Bluetooth device in the hearing range is an indicator of contact. This is an easy-to-use approach that unfortunately does not do well on the accuracy front. After all, the range of Bluetooth is around 100 feet. So, if you go to a large retail store, everyone in the store might end up being identified as a contact (large false-positive rate). To alleviate this concern, the second level of this tracing is based on signal strength. When the signal strength crosses a threshold, i.e. the device is not just in the hearing range but also close enough to deliver a strong signal, it is identified as a contact.
This is the most common approach out there today and it forms the bedrock of the Apple-Google Exposure Notification system. However, as the founders of Bluetooth have themselves pointed out, this approach is bound to have false-positives and false-negatives as well. The premise of this approach is that the signal strength is a function of distance, therefore signal strength can be used as a proxy measure for distance. Unfortunately, the signal strength depends on other factors too like if the phone is in a user’s hand or pocket or if there is an obstacle between the two devices. This leads to large errors in distance estimation. Two devices 3 feet away can appear to be 30 feet away and vice-versa. In fact, such inaccuracies in Bluetooth positioning are common knowledge among researchers in the wireless positioning community. However, in the absence of more advanced techniques that are as ubiquitous as Bluetooth, this method has emerged as the method of choice.
Another obvious solution to proximity detection is GPS. We can use GPS to track every smartphone and identify two phones as being in contact when they stay too close for too long. The premise is promising: after all, most smartphones are GPS-enabled, GPS works across the globe. So, a GPS-based solution would be ubiquitous and scalable.
However, this approach fails on two axes. Not only does this raise serious privacy questions (you are always being tracked), it fails on the accuracy front as well. GPS suffers from poor accuracy, especially indoors. Research has shown that it is non-trivial to enable robust GPS indoors and even when it works, it achieves positioning errors of around 10 meters (nearly 30 feet). These errors are too high for contact-tracing approaches. Recall that a contact tracing approach needs to identify if two people are closer than 6 feet (2 meters) of each other. An error of 30 feet will throw these estimates way off and will lead to high false-positive or false-negative rate, depending on how the system is designed.
Recording GPS traces of every individual all the time also raises obvious privacy concerns. GPS traces are easy to de-anonymize and can reveal so much about individuals. I must note that contact tracing advocates believe that these concerns are overshadowed by the health benefits of the contact tracing apps. After all, they claim, we share our location data with corporations every day when we check their apps for maps, ride-hailing, or food delivery. I believe the key element here is control. As a user, you have the choice to not use these apps — this has happened in the past when folks mass deleted certain apps due to privacy violations. However, if these apps are made compulsory by employers or governments, the use of GPS is no longer optional and is a recipe for gross privacy violations.
Overall, the accuracy of GPS is low for contact tracing, so using it for contact-tracing won’t help much anyway. However, with explicit user consent to upload selective data, it can provide other benefits like helping governments visualize where high-risk individuals are and to help allocate healthcare resources, thereby preempting the outbreak.
Sound is as ubiquitous as Bluetooth. After all, the primary purpose of phones is to emit, record, and transfer sound (or it used to be). Thus, sound-based range estimation is another competing approach. The overall idea is similar to Bluetooth-based contact tracing. Every device emits some random but unique sound signature. Everyone who listens to that sound signature can use it as a signal to infer distance. Similar to Bluetooth, you can use the loudness of the sound. Of course, the sound is emitted at low amplitude and on frequencies outside the human hearing range, so that it does not annoy humans.
Unlike Bluetooth and GPS, sound is a mechanical wave that travels much slower than radio waves (a million times slower). Therefore, acoustic-ranging has another arrow in its quiver: time-of-flight estimation. Specifically, two devices can accurately measure the time it takes for a sound to go from one device to another. Since sound travels as a wave, we can multiply this time by the speed of sound to get the distance. For example, if it took a sound signal 4 milliseconds to go from my phone to your phone, it meant the phones are at a distance of about 4.5 feet. Even if you make an error of a millisecond or two in measuring this time, the errors won’t be large. More importantly, the time estimates are more robust to reflections and obstacles than strength estimates.
At a high level, this idea is very promising: it is accurate and ubiquitous. However, as they say, there are no free lunches. Sound-based ranging has three key pitfalls:
Scale: The number of devices that can simultaneously perform sound-based ranging is small. It works fine when one device talks to another, but as the number of devices talking to each other increases, it becomes very noisy. Add to it the ambient noise in places like retail stores and you have a very flaky system in your hands. I believe new ideas are required (and being developed) to solve this challenge in sound-based ranging.
Privacy: Imagine asking people to keep the microphones on their phones turned on all the time. Your phone will hear everything you say. This is a level of privacy-risk that most people are unwilling to tolerate.
Animal discomfort: Even though humans cannot hear these minute sounds transmitted at frequencies outside human hearing range, animals can. A lot of tests will be required to ensure that we do not end up with scenarios like the finale sequence from Silicon Valley (spoiler alert!) where rats are enticed by human-inaudible sounds from phones.
The methods I talked about so far are relatively mainstream. There are several other interesting ideas floating around in the world of research. Here are a few:
Time-of-flight using Bluetooth: Can we use time-of-flight measurements in Bluetooth like we do using sound waves? This is a very interesting premise. It will alleviate the concerns about sound-based ranging and will remove inaccuracies of Bluetooth contact-tracing. The challenge is that any such measurements will need to be way more accurate. Radio waves travel a million times faster than sound, so we need to be a million times more accurate in our time measurements. In fact, I wrote two papers about how we could achieve a nanosecond-level accuracy in time measurements and achieve a foot-level accuracy in distance measurements by using advanced phase information. Alas, phase information isn’t available on all smartphones today, so this approach, while way more accurate, won’t be ubiquitous for the next few years.
Hybrid multi-modal approaches: Many app designers today are exploring the use of different modalities like sound and Bluetooth in conjunction with each other, so that one technology can help remove the pitfalls of the other. Such approaches are seeing a lot of traction and are being used in many of the apps out there.
I’ll sign off with two key takeaways. First, the fact that all the technologies out there have some pitfalls today does not mean that we shouldn’t use any of them. We should be aware of the trade-offs that exist and make our choices based on the objective function that we intend to optimize. Moreover, there is research being done today to alleviate these pitfalls. Second, this presents an opportunity for us to develop new tools that can be better than any of the ones we have at our disposal. My goal with this set of blog posts has been to add more nuance and information to the contact tracing debate. Hopefully, this will help us make better decisions about our future course of action. Feel free to reach out if you need more answers about any of the technologies mentioned here.
Originally published at https://www.linkedin.com.