BATL

BATL: Biometric Authentication with a Timeless Learner

Funding Agency: US Intelligence Advanced Research Projects Activity (IARPA)

Lead organization: USC’s Information Sciences Institute

Team members: USC’s Viterbi School of Engineering Computer Science Department, Idiap Research Institute, Hochschule Darmstadt and Norwegian University of Science and Technology, TREX Enterprises, Northrop Grumman Corporation Grumman Corporation

Performance period: March 2017 – February 2021

Project description: Within the digital world in which access to enterprise, personal, and societal information increasingly relies upon automatic authentication systems, the goal of IARPA’s Thor program is to utilise Presentation Attack Detection (PAD) methods to identify known and unknown Presentation Attacks (PA) in a biometric recognition system. A biometric PA, also commonly referred to as biometric spoofing, is a method which inhibits the intended operation of a biometric capture system, interfering with the recording of the true sample/identity, ultimately preventing the subject from being correctly identified. Typical PAs utilise an artefact to present an alternative biometric characteristic of the victim (i.e., targeted impostor attack).

The USC Information Sciences Institute (USC ISI)-led team will develop an end-to-end multimodal PAD system to deliver robust, accurate, and timely detection of known and unknown PAs, through a novel synthesis of (i) new and existing sensors, and (ii) machine learning techniques. BATL’s sensor suite includes conventional and novel, unconventional face, iris, and fingerprint sensors that, together, provide a rich set of signals for PAD, while producing data that enable interoperability with existing COTS comparison algorithms.

In particular, the BATL system includes three modality-specific PAD modules (i.e., face, iris and fingerprint), each of which extracts various types of features from the input. Decisions from all modality-specific PAD modules as well as an additional unknown attack detector are all fused to produce a robust decision to discriminate bona fide presentation from PAs. We will also make a distinction between impostors and identity concealers.

The biometric research group at Hochschule Darmstadt in cooperation with the Norwegian Biometrics Laboratory (NTNU) participates in the following tasks:

  • Presentation Attack Detection for Fingerprint
  • Presentation Attack Detection for Iris