Search

Search Results

• METHODS OF IMPROVED USER IDENTIFICATION BASED ON LIVENESS DETECTION TECHNOLOGY

  V.V. Zolotarev, А. О. Povazhnyuk, Е.А. Maro

  2022-05-26

  Abstract ▼

  Biometric identification and access control systems contain methods for recognizing a subject's
  personality based on his unique physiological and behavioral characteristics. The purpose of
  this work is to develop a system for secure interaction (authentication) of participants in gamified
  educational projects, which includes countering security threats that arise when using biometric user
  characteristics. A comparative analysis of the efficiency of recognition of fake biometric samples by
  liveness detection methods based on the detection of sample substitution using a photo, video on a
  display, a 3D model, and a mask has been performed. During research a method of using the liveness
  detection for include to a gamified educational environment system was proposed. A modification of
  the liveness detection method (hybrid method) has been proposed and a biometric identification system
  in real time has been designed using the proposed method. A two-stage hybrid biometric identification
  method has been developed based on the joint use of passive and active software methods for
  detecting fake biometric samples. The method is adapted for use with a minimum number of additional
  devices, the only biometric feature scanner is a 2D-camera. The network of types two-layer
  perceptron, three-layer perceptron and convolutional neural network was tested. The network was
  trained on the author's training examples. The position of the announcer when recording training
  examples: the distance of the face from the camera is 60cm, the recording modes when the head is
  turned by 0 (look directly into the camera), 30 (the head is slightly turned to the side) and 45 (the
  head is turned strongly to the side) degrees. Based on the testing results, the best recognition rates
  were found in a convolutional neural network with 3 convolutional layers and 1 fully connected one.
  Accuracy of recognition of the spoken word is obtained up to 100% when the user's head is turned up
  to 30° and up to 70% - when the user's head is turned up to 45°. The FAR value of this system was
  1%, the FRR value was 0% for testing on 1000 samples.