The prevalence of smartphones have changed our lives, perhaps mostly for better but in a few cases they have caused harm. Gig-economies only made possible thanks to smartphones, and interestingly enough there are still disputes on whether they cause more good or harm, or differently put, what is their net contribution to the society.
In case you have been living under rock. gig-economy usually refers to the work that is distributed on-demand to temporary workers through a website or app (see [1] for more); well-known examples are Amazon Mechanical Turk for crowd-work, deliveroo in food delivery and the bigger players such as Uber, Lyft or Bolt that provide transportation on-demand.

Ride-sharing apps have revolutionized transportation, but they have their own problems; There have been many reports of sexual assault and various other crimes by ride-sharing drivers [2], in addition there are cases that drivers commit fraud or game the system. They create fake GPS traces, create ride requests from stolen accounts and more [3]. One way of tackling some of these issues is performing background checks by the ride-sharing companies, which is mandatory in certain jurisdictions. This has reduced the crime rates already but how about stolen accounts/phones or when a group a drivers share the same account (assuming to avoid a background check). In such cases, your Uber driver may not be who you think they are. Uber has a system in place that sometimes asks the driver to stop and authenticate themselves using a selfie picture [4]. It is unclear how often or under what criteria they trigger this. But would not it be more efficient if they could analyse the driving style of the driver and confirm the driver’s identity based on that? This is where driver identification comes into play.

Driver identification is a term used by the research community that refers to methods that identify the driver based on their driving behavior, think of it an authentication method, like a password, or Touch ID, but it is used more like your 2-factor authentication system, it can be used as a secondary way of authentication that may trigger a more reliable authentication method. What is driver behavior? Driver behavior is anything that you do while driving, but most current research are focused on sensor data that can be collected from the car itself, generally through OBD port or CAN bus. It is also possible though to construct a model only from GPS data from driver’s smartphone.

This technology can be used to verify ride-sharing drivers’ identity. What are other applications of such a system? Obviously mass surveillance! Just kidding, but that is something to be concerned about. Freight companies need to verify that legitimate driver is behind the wheel, also long haul drivers are subject to daily and weekly working limits, driver identification can help detecting cheating drivers and even companies. You can achieve the same goal by installing a camera in the vehicle, but that is more costly and not many people would be happy about being watched all the time. Insurance companies would be able to prevent fraud, for example in cases that the contract is meant to cover only a particular person. Forensics is another application area, sometimes there are disputes about who was actually driving the vehicle period to crash. There are many other use-cases and the ones that I have not thought of yet!

For the past 4 years, driver identification has been my main research topic and the two applications I am excited to see such a solution be applied to is for ride-sharing drivers to ensure safety and peace of mind of passengers, and long haul drivers, because an overworked and sleep deprived driver of a heavy vehicle not only risks their own life they are also extremely dangerous to other road users. Hopefully soon self-driving cars will take over and make these solutions irrelevant but we have a long journey ahead and we will not get there anytime soon.

But how is it possible?

The power of AI, it can solve any problem! In my research work I have explored many approaches and depending on the application and available data one can choose the most suitable method. I started with traditional machine learning models, and ended up using deep learning based methods, but there is no right answer, it all depends on the available data and requirements.
Generally there are two cases:

1. You have access to frequent data from car, often CAN-bus. In this case we use pattern recognition to build a profile for the driver and this profile can later be used to classify or verify driver’s identity. Frequent data (~20Hz and above) often leads to high accuracies.
2. You only have access to infrequent (~1Hz) location data. We still can do a lot with this data, every driver take hundreds of decisions when driving and some of these decisions are unique to a certain degree. For example, choice of speed, acceleration pattern, speeding and its frequency all these have been shown in our research that can be sufficient to be able to discriminate drivers with good accuracy but detection will be slower.

If you like to learn more or exchange ideas about driver identification please drop me an email or write me on Twitter! You can find my relevant publications here.
Also I am looking for the next challenge in my career either in industry or academia.

References

1. V. De Stefano, ‘The Rise of the Just-in-Time Workforce: On-Demand Work, Crowdwork, and Labor Protection in the Gig-Economy’, Comp. Lab. L. & Pol’y J., vol. 37, p. 471, 2016 2015.
2. Who’s driving you?
3. Uber fights off scammers every day. Here’s how it learned the tricks
4. Engineering Safety with Uber’s Real-Time ID Check