TY - GEN
T1 - Temporal Convolutional Networks for Robust Face Liveness Detection
AU - Padnevych, Ruslan
AU - Carmo, David
AU - Semedo, David
AU - Magalhães, João
N1 - Publisher Copyright:
© 2022, Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - Face authentication and biometrics are becoming a commodity in many situations of our society. As its application becomes widespread, vulnerability to attacks becomes a challenge that needs to be tackled. In this paper, we propose a non-intrusive on the fly liveness detection system, based on 1D convolutional neural networks, that given pulse signals estimated through skin color variation from face videos, classify each signal as genuine or as an attack. We assess how fundamentally different approaches – sequence and non-sequence modelling – perform in detecting presentation attacks through liveness detection. For this, we leverage on the Temporal Convolutional Network (TCN) architecture, and exploit distinct and TCN grounded types of convolution and architectural design schemes. Experiments show that our TCN model provides the best balance in terms of usability and attack detection performance, achieving up to 90% AUC. We further verify that while our 1D-CNN with a residual block variant performs on par with the TCN model in detecting fake pulses, it underperforms in detecting genuine ones, leading to the conclusion that the TCN model is the most adequate for a production environment. The dataset will be made publicly available to foster research on the topic. (https://github.com/novasearch/Mobile-1-D-Face-Liveness-Detection
AB - Face authentication and biometrics are becoming a commodity in many situations of our society. As its application becomes widespread, vulnerability to attacks becomes a challenge that needs to be tackled. In this paper, we propose a non-intrusive on the fly liveness detection system, based on 1D convolutional neural networks, that given pulse signals estimated through skin color variation from face videos, classify each signal as genuine or as an attack. We assess how fundamentally different approaches – sequence and non-sequence modelling – perform in detecting presentation attacks through liveness detection. For this, we leverage on the Temporal Convolutional Network (TCN) architecture, and exploit distinct and TCN grounded types of convolution and architectural design schemes. Experiments show that our TCN model provides the best balance in terms of usability and attack detection performance, achieving up to 90% AUC. We further verify that while our 1D-CNN with a residual block variant performs on par with the TCN model in detecting fake pulses, it underperforms in detecting genuine ones, leading to the conclusion that the TCN model is the most adequate for a production environment. The dataset will be made publicly available to foster research on the topic. (https://github.com/novasearch/Mobile-1-D-Face-Liveness-Detection
KW - 1D convolutional neural networks
KW - Liveness detection
KW - Mobile presentation attacks
KW - Temporal convolutional networks
UR - http://www.scopus.com/inward/record.url?scp=85129874158&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-04881-4_21
DO - 10.1007/978-3-031-04881-4_21
M3 - Conference contribution
AN - SCOPUS:85129874158
SN - 978-3-031-04880-7
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 255
EP - 267
BT - Pattern Recognition and Image Analysis - 10th Iberian Conference, IbPRIA 2022, Proceedings
A2 - Pinho, Armando J.
A2 - Georgieva, Petia
A2 - Teixeira, Luís F.
A2 - Sánchez, Joan Andreu
PB - Springer
T2 - 10th Iberian Conference on Pattern Recognition and Image Analysis, IbPRIA 2022
Y2 - 4 May 2022 through 6 May 2022
ER -