Swiss team claims to make stripped-down biometric presentation attack detection for cars
Researchers at a Swiss institute say they have created a highly reliable one-camera way to detect face presentation attacks in vehicles that require continuous biometric verification of a driver’s identity.
Currently, some vehicle makers are installing facial recognition systems, some of which are limited to one-time verification in order to start the engine.
In some of those cases, according to scientists at the independent, non-profit Idiap Research Institute, a presentation attack as rudimentary as holding up a tablet with the photo of the authorized operator is enough to circumvent security.
The Idiap team’s goal was to build a near-real time verification and biometric liveness product that operates in the resource-constrained environment of a vehicle. That meant, among other considerations, that a successful system had to require little memory and limited sensing abilities while delivering rapid, continuous inference.
In a new paper, the researchers claim their system uses just one full-high definition near-infrared (940-nm) camera on the front end with a “lightweight” face presentation attack detection, or face PAD, framework based on a nine-layer convolutional neural network that achieved 98 percent overall accuracy.
Not unexpectedly given the setting, the researchers had to deal with a paucity of training near-infrared face PAD data. To get around that, they say combined the adaptation of domain-specific layers and the task fine-tuning of a base convolutional neural network.
The Idiap team gathered a face PAD dataset with more than 5,800 near-infrared videos called the ‘in-Vehicle Face Presentation Attack Detection’ dataset.
More than 4,000 bona-fide recordings of 40 subjects (24 male, 16 female of “various” ethnicities) are in the dataset.
So are almost 1,800 attack videos from 89 presentation attack instruments, including rigid and flexible masks, photo prints and digital displays for replay attacks.
Apple, still rumored to be working on an autonomous car, is already working on upgrades to the face biometrics attack detection it has embedded in the iPhone.
The list of other vendors developing in-vehicle driver monitoring products is a long one, and include LG Electronics, Cerence, Mitsubishi and Cipia.
automotive biometrics | biometric liveness detection | biometrics | biometrics at the edge | face biometrics | facial recognition | Idiap | presentation attack detection | spoof detection