Contributed by Bob Chabot
Fingerprinting Driver Behavior
Welcome to extra sensory suspension
Modern vehicles are smart enough to witness driving behavior, assist vehicle service professionals and help improve the driving experience for motorists. So imagine if individual driving styles could be identified and "fingerprinted" with 100 percent accuracy, based only on hard data collected from in-vehicle sensor data on vehicle Controller Area Networks (CAN). The good and bad news: It's already been done, with 100 percent accuracy. But there's a catch.
Automobiles Have Evolved Beyond Purely Mechanical Roots
Recently, a team of researchers from the University of Washington and the University of California at San Diego, led by Miro Enev, a machine-learning engineer at Belkin Corp., found that individual driving behavior was not only as unique as fingerprints, it could also be reliably identified. "With very limited amounts of driving data we can determine very powerful and accurate inferences about the driver's identity," Enev explained. The team's research paper, Automobile Driver Fingerprinting, was recently presented at the July 2016 Privacy Enhancing Technology Symposium in Germany.
"By logging and analyzing the data streams from 16 sensors that broadcast over the CAN system while vehicles were being driven, we found time to be the determinant for accurately fingerprinting the driving style of specific individuals. In just 15 minutes of operation using only the brake pedal position sensor to collect driver behavior data, we were able to accurately identify individual drivers with 87 percent accuracy. When we used all of the sensors and monitored driving for just 15 minutes, we achieved 100 percent accuracy — every single time."
The implications of driver fingerprinting span the spectrum from good to bad, according to Enev. Some are positive, such as vehicle theft detection, improved vehicle maintenance or expanding the range of useful aftermarket services to motorists. On the other hand, concerns about potential conflicts-of-interest, personal privacy breaches, authorizing access and other issues exist and need resolution.
"The list of 16 sensors used by Enev's team are present in most modern vehicles and provide a baseline from which to measure and record driver fingerprinting data. The values produced by this sensor subset are dependent on driver behavior, rather than on the behaviors of internal vehicle systems." (All images — Automobile Driver Fingerprinting, by Miro Enev et al).
New Opportunities are Emerging
Modern vehicles are intelligent cyberphysical platforms built on sophisticated sensors, powerful embedded computers, telematics and connected car technologies. Collectively, they have enabled mobile communication and multi-way data streaming with automakers, service facilities, other third parties (e.g. insurance firms, Google, Apple) and entire transportation system infrastructures.
"These new technologies are unlocking new levels of innovation in the automotive marketplace aimed at improving serviceability, safety, efficiency, engagement and ultimately, the driving experience," Enev noted. "In addition, the potential to monetize vehicle data streams and the ubiquity of connectivity options has fostered the emergence and growth of a novel data market."
He cited several examples within the rapidly growing data-sharing aftermarket economy:
- Usage-based insurance dongles — Many insurance companies offer "pay as you drive" discounts to enable rate reductions for consistently safe driving behaviors. Drivers opt-in by attaching a dongle (with built in GPS and telematics) to the vehicle OBD-II port, which analyzes and transmits the data for upstream processing (e.g. Progressive's Snapshot, State Farm's In-Drive).
- Automotive dongles — Tier One suppliers and other automotive aftermarket manufacturers have developed and sold products that offer diagnostics and connected shop solutions (e.g. Delphi Automotive), stop-start functionality (e.g. Voyomotive LLC), and other functionality (e.g. commercial fleet monitoring).
- Concierge services — Many new startups, such as Automatic and Zubie, can turn any car into a "smart car" by using a relatively cheap device (approximately $100 or less) that beams data from the vehicle's internal network to the driver's phone via dongle or Bluetooth. Apps then allow drivers to gain start-stop ability, recall where they parked, alerting carpool/friends about arrival times, gamification (e.g. break your fuel efficiency record for home commutes) and more.
Balancing Utility with Privacy
The technologies allow any OBD-II equipped vehicle manufactured after 1995 to be an active participant in the driver's technology ecosystem, enable interactions with other smart devices and apps, and provide enriched driver experiences. But many of the dongle features and applications enabled by providers require making and using judgments that are not always made public upfront. Tensions regarding data ownership, access to it, and personal privacy are increasing.
Think about servicing vehicles, where driving conditions and style can be crucial information to diagnose and remedy problems, such as driveability or seemingly premature wear issues. We ask for it, but can we really trust it? To date, that information has been typically anecdotal in nature. A technician or service advisor asks the customer to describe what's happening, how and where they drive, etc. But like doctors often find with their patients, the information shared is frequently incomplete and less than accurate.
"Combining customer anecdotes with empirical data enhances diagnostics and facilitates more appropriate remedies," Enev suggested. "Accessing readily available empirical vehicle data would provide trustable information that service professionals could use in conjunction with the driver's story to provide more complete and driver-appropriate solutions."
Sensors can be clustered (left) into four groups: Acceleration (accelerator pedal, torque requested, etc.), denoted in blue; Turning (steering angle, lateral acceleration, etc.), denoted in red; Vehicle State (velocity, gear, RPM, etc.), denoted in green; and Deceleration (brake pedal, maximum achievable torque, etc.) denoted in purple. Enev et al were able to determine and rank which sensors were better able to differentiate between drivers and show their relative reliability as identifiers (right). The top five sensors are shown in red.
Does Privacy Matter Enough?
"Not many vehicle owners think about the risks they are exposed to," the research team noted. "Instead they're mostly just giving this data from their vehicle to automakers and other parties, which points to a more fundamental problem with automotive security: There should be a permission structure built around every sensor stream."
"Repercussions, with or without upfront driver authorization, have already occurred; without adequate safeguards, more are likely," suggested Enev. "in the past, automakers, law enforcement, lawyers, insurance companies, giant tech companies, and other third parties have successfully subpoenaed and/or used vehicle data in legal proceedings —ranging across criminal acts, infidelity, insurance fraud, and the harvesting of personal data for marketing purposes.
The research paper cited two examples from the automotive world. Tesla Motors CEO Elon Musk recently used vehicle sensor data in court, to dispute the claim of a New York Times journalist about the limited range of his car's electric batteries. Using vehicle data the journalist had approved the use of when the vehicle was purchased, Musk demonstrated the journalist, contrary to the details of his claim, had taken an undisclosed detour and did not fully charge his vehicle.
To demonstrate how readily available and easily extracted vehicle data is, Jim Farley, an executive vice president for Ford Motor Co., has stated: "We know everyone who breaks the law. We know where and when you are doing it. We have GPS and other technologies in your car, so we know what you are doing."
Sensor data provides a witness to individual driver behaviors when operating vehicles.
The top image shows two consistent behaviors of this driver: (1) The high amount of steering wheel activity required (black curve) and its close correlation with the lateral acceleration measurements (gray curve), and (2) The amount of braking (red curve) in the early part of turns and the subsequent accelerations (green curve) during exits.
The lower image isolates just the velocity data shown throughout the road course used by the research team. Note the difference in acceleration and deceleration technical driving patterns across the different road segments, as highlighted with boxes of different colors (i.e., interstate = green, urban = pink).
Build-In Permission-Based Security Measures
"The industry must safeguard those instances where the vehicle owner didn't approve access, or did not fully understand what that access entailed," Enev claimed. "It's doable."
He then suggested three criteria that could serve as prerequisite guidelines for any user wanting access to any individual's vehicle data:
- Prove Need-to-Know — Every new application that exposes a vehicle owner's data to a threat — malicious, adversarial or otherwise — should be approached on a need-to-know basis. For instance, a gadget meant to track your fuel efficiency doesn't need and shouldn't be able to track every push of your brake pedal or turn of the steering wheel.
- Full Explanations by Users — How data and information will be harvested and how it will be used should be completely divulged to vehicle owners upfront.
- Obtain Written Permission — Instead of making any or all of a vehicle's data and sensor systems readily and easily available to any device connected to the CAN bus, written permission systems should be in place beforehand, just as the computer industry does with iOS and Android operating systems.
Driver Fingerprinting Can Serve Everyone's Interests
"The diversity of existing legal opinions on other issues in the automotive world — right-to-repair, vehicle software copyrights, and others — highlights the complexity of creating regulatory frameworks for intelligent automotive systems," Enev advised.
From a legal perspective today, he noted there are varying stances on vehicle sensor data ownership, processing and management. Within the United States, 13 states have now adopted the stance that a vehicle's sensor data is private and the property of the vehicle owner. However, within these 13 states, there are marked differences on what constitutes acceptable data retrieval without owner consent.
"We hope that our efforts help guide the design of policy, safeguards and mechanisms to balance the utility and privacy tensions emerging in modern automotive contexts'" Enev stated. "In an ever-connected and automated landscape, our work is relevant and can provide substantial benefits to automobile manufacturers, suppliers, service and repair professionals, drivers, and other participating stakeholders in the aftermarket. Done right, data access, utility and privacy can co-exist."
[Editor's note: For the latest diagnostic and automotive service insights, read MOTOR Magazine's July 2016 issue.]