您的位置:山东大学 -> 科技期刊社 -> 《山东大学学报(工学版)》

山东大学学报 (工学版) ›› 2023, Vol. 53 ›› Issue (6): 1-7.doi: 10.6040/j.issn.1672-3961.0.2023.132

• 机器学习与数据挖掘 •    下一篇

一种高效且轻量的RGB单帧人脸反欺诈模型

李家春(),李博文,常建波   

  1. 华南理工大学计算机科学与工程学院,广东 广州 510006
  • 收稿日期:2023-06-21 出版日期:2023-12-20 发布日期:2023-12-19
  • 作者简介:李家春(1968—),女,湖北松滋人,副教授,硕士生导师,博士,主要研究方向为计算机网络与信息安全、隐私保护、人工智能安全、智慧教学等。E-mail: jclee@scut.edu.cn
  • 基金资助:
    教育部产学合作协同育人资助项目(201902186007);教育部产学合作协同育人资助项目(201901034001)

An efficient and lightweight RGB frame-level face anti-spoofing model

Jiachun LI(),Bowen LI,Jianbo CHANG   

  1. School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, Guangdong, China
  • Received:2023-06-21 Online:2023-12-20 Published:2023-12-19

摘要:

针对在仅具有三原色(red-green-blue,RGB)摄像头的通用消费设备上部署基于深度学习的人脸反欺诈(face anti-spoofing,FAS) 算法时存在的挑战问题,提出一种高效且轻量的RGB单帧FAS(efficient and lightweight RGB frame-level face anti-spoofing,EL-FAS)模型。探索一种新的全局空间自注意力机制捕获全局上下文信息的依赖关系,以提高模型泛化能力并在受限条件下实现高检测性能;设计一种等通道像素级二元监督方法,强制模型从不同的像素中学习共享特征;采用Bottleneck模块搭建骨干网络以减少模型参数。试验结果表明,EL-FAS模型在OULU-NPU数据集的大多数协议上平均分类错误率RACE最低,取得较好的人脸欺诈检测效果,在SiW数据集和跨数据集测试中也取得较好的性能,并且模型轻量,参数只有1.34×106个。

关键词: 深度学习, 人脸反欺诈, 自注意力机制, 像素级监督, 轻量级模型

Abstract:

Based on the challenge when deploying a deep learning-based face anti-spoofing (FAS) algorithm on general-purpose consumer devices with only RGB camera, an efficient and lightweight RGB frame-level FAS model (EL-FAS) was proposed. To improve the generalization ability of the model and achieve high performance under constrained conditions, a novel global spatial self-attention mechanism was explored to capture global feature dependencies, and an equal-channel pixel-wise binary supervision method was designed to force our model to learn shared features from different pixels. The Bottlenecks residual block was used to establish the backbone network to reduce parameters. Analysis and the experimental results showed that EL-FAS model achieved state-of-the-art performance in the OULU-NPU dataset, obtained competitive performance in the SiW dataset and cross-dataset tests. The model was lightweight, with only 1.34×106 parameters.

Key words: deep learning, face anti-spoofing, self-attention mechanism, pixel-wise supervision, lightweight model

中图分类号: 

  • TP391

图1

EL-FAS模型架构"

表1

EL-FAS模型架构细节"

解编码器 模块 输入 操作 t c s
编码器 E1 3×2242 Conv3×3 32 2
32×1122 Bottleneck 1 16 1
E2 16×562 Bottleneck 6 24 2
24×562 Bottleneck 6 24 1
24×562 Bottleneck 6 32 2
E3 32×282 Bottleneck 6 32 1
32×282 Bottleneck 6 32 1
32×282 Bottleneck 6 64 2
64×142 Bottleneck 6 64 1
64×142 Bottleneck 6 64 1
E4 64×142 Bottleneck 6 64 1
64×142 Bottleneck 6 96 1
96×142 Bottleneck 6 96 1
96×142 Bottleneck 6 96 1
解码器 96×142 ConvTrans 32 2
D1 64×282 Bottleneck 6 32 1
32×282 Bottleneck 6 32 1
32×282 ConvTrans 24 2
D2 48×562 Bottleneck 6 24 1
24×562 Bottleneck 6 24 1
24×562 ConvTrans 16 2
D3 32×1122 Bottleneck 6 16 1
16×1122 Bottleneck 6 16 1
D4 16×1122 ConvTrans 16 2
16×2242 Conv3×3 3 1

图2

全局空间自注意力机制架构"

表2

数据集描述"

数据集 用户数 视频数 攻击设备 测试协议数
SiW 165 4 478 iPad, iPhone7, GalaxyS8, Asus MB168B(RGB) 3
OULU-NPU 55 4 950 Dell 1905FP, MacBook(RGB) 4
Replay-Attack 50 1 200 iPad, iPhone 3GS(RGB) 1
CASIA-FASD 50 600 A4 Paper, iPad(RGB) 1

表3

数据集内消融试验结果"

模型 RAPCE RBPCE RACE
Single 4.2±9.3 4.6±6.8 4.4±4.9
Equal 3.3±4.7 5.4±6.8 4.4±3.2
Single+Attention 3.4±4.5 4.4±6.8 3.9±3.5
Equal+Attention 1.6±3.7 5.8±8.3 3.7±4.0

表4

跨数据集消融试验结果"

模型 RAPCE RBPCE RACE
Single 3.3±3.7 2.5±2.5 2.9±2.7
Equal 3.0±4.0 2.5±2.2 2.8±2.0
Single+Attention 3.0±4.0 2.0±1.9 2.5±1.8
Equal+Attention 0.8±1.8 3.7±4.3 2.3±2.4

表5

OULU-NPU数据集内的测试结果"

协议 方法 RAPCE RBPCE RACE
1 GRADIANT[18] 1.3 12.5 6.9
Auxiliary[17] 1.6 1.6 1.6
FaceDs[19] 1.2 1.7 1.5
STDN[20] 0.8 1.3 1.1
DSGT[5] 2.0 0 1.0
DeepPix[10] 0.8 0 0.4
EL-FAS 0 1.6 0.8
2 DeepPix[10] 11.4 0.6 6.0
FaceDs[19] 4.2 4.4 4.3
Auxiliary[17] 2.7 2.7 2.7
GRADIANT[18] 3.1 1.9 2.5
DSGT[5] 2.5 1.3 1.9
STDN[20] 2.3 1.6 1.9
EL-FAS 0.3 0.5 0.4
3 DeepPix[10] 11.7±19.6 10.6±14.1 11.1±9.4
GRADIANT[18] 2.6±3.9 5.0±5.3 3.8±2.4
Auxiliary[17] 2.6±3.9 5.0±5.3 3.8±2.4
FaceDs[19] 4.0±1.8 3.8±1.2 3.6±1.6
STDN[20] 1.6±1.6 4.0±5.4 2.8±3.3
DSGT[5] 3.2±2.0 2.2±1.4 2.7±0.6
EL-FAS 1.9±2.4 3.5±2.4 2.7±1.2
4 DeepPix[10] 36.7±29.7 13.3±16.8 25.0±12.7
GRADIANT[18] 5.0±4.5 15.0±7.1 10.0±5.0
Auxiliary[17] 9.3±5.6 10.4±6.0 9.5±6.0
FaceDs[19] 5.1±6.3 6.1±5.1 5.6±5.7
DSGT[5] 6.7±7.5 3.3±4.1 5.0±2.2
STDN[20] 2.3±3.6 5.2±5.4 3.8±4.2
EL-FAS 1.6±3.7 5.8±8.3 3.7±4.0

表6

SiW数据集内的测试结果"

协议 方法 RAPCE RBPCE RACE
1 Auxiliary[17] 3.58 3.58 3.58
DSGT[5] 0.64 0.17 0.40
STDN[20] 0 0 0
EL-FAS 0 0 0
2 Auxiliary[17] 0.57±0.69 0.57±0.69 0.57±0.69
DSGT[5] 0 0.04±0.08 0.02±0.04
STDN[20] 0 0 0
EL-FAS 0.08±0.14 0 0.04±0.07
3 Auxiliary[17] 8.31±3.81 8.31±3.80 8.31±3.81
STDN[20] 8.33±3.33 7.50±3.33 7.90±3.30
DSGT[5] 2.63±3.72 2.92±3.42 2.78±3.57
EL-FAS 25.80±23.50 0 12.90±11.70

表7

CASIA-FASD和Replay-Attack数据集间的跨数据集测试结果"

方法 RHTE
CR RC
CNN[3] 48.5 45.5
FaceDs[19] 28.5 41.1
Auxiliary[16] 27.6 28.4
EL-FAS 31.2 31.1

表8

从SiW到OULU-NPU数据集的测试结果"

协议 方法 RAPCE RBPCE RACE
1 Auxiliary[17] 10.0
DSGT[5] 1.7 13.3 7.5
EL-FAS 3.3 4.6 3.9
2 Auxiliary[17] 14.1
DSGT[5] 9.7 14.2 11.9
EL-FAS 3.1 20.5 11.8
3 Auxiliary[17] 13.8±5.7
DSGT[5] 17.5±4.6 11.7±12.0 14.6±4.8
EL-FAS 3.9±2.8 10.8±3.8 7.4±2.5
4 Auxiliary[17] 10.0±8.8
DSGT[5] 0.8±1.9 10.0±11.6 5.4±5.7
EL-FAS 0.8±1.8 3.7±4.3 2.3±2.4

表9

轻量级指标对比"

方法 参数量 浮点运算量
LSTM-CNN[4] 2.07×106 4.77×109
DSGT[5] 5.53×106 773.40×109
Depth[8] 9.25×106 20.57×109
BASN[9] 80.40×106 274.13×109
DeepPix[10] 3.20×106 4.62×109
EL-FAS 1.34×106 0.79×109
1 PATEL K , HAN H , JAIN A K . Secure face unlock: spoof detection on smartphones[J]. IEEE Transactions on Information Forensics and Security, 2016, 11 (10): 2268- 2283.
doi: 10.1109/TIFS.2016.2578288
2 KOMULAINEN J, HADID A, PIETIKÄINEN M. Context based face anti-spoofing[C]//Proceedings of the 2013 IEEE Sixth International Conference on Biometrics: Theory, Applications and Systems (BTAS). Arlington, USA: IEEE, 2013: 6712690.
3 YANG J, LEI Z, LI S Z. Learn convolutional neural network for face anti-spoofing[EB/OL]. (2014-08-26)[2023-02-21]. https://arxiv.org/abs/1408.5601.
4 XU Z, LI S, DENG W. Learning temporal features using LSTM-CNN architecture for face anti-spoofing[C]//Proceedings of the 2015 3rd IAPR Asian Conference on Pattern Recognition(ACPR). Kuala Lumpur, Malaysia: IEEE, 2015: 141-145.
5 WANG Z, YU Z, ZHAO C, et al. Deep spatial gradient and temporal depth learning for face anti-spoofing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, USA: IEEE, 2020: 5042-5051.
6 ZHANG S, WANG X, LIU A, et al. A dataset and benchmark for large-scale multi-modal face anti-spoofing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach, USA: IEEE, 2019: 919-928.
7 WANG Z , WANG Q , DENG W , et al. Learning multi-granularity temporal characteristics for face anti-spoofing[J]. IEEE Transactions on Information Forensics and Security, 2022, 17, 1254- 1269.
doi: 10.1109/TIFS.2022.3158062
8 ATOUM Y, LIU Y, JOURABLOO A, et al. Face anti-spoofing using patch and depth-based CNNs[C]//Proceedings of the 2017 IEEE International Joint Conference on Biometrics (IJCB). Denver, USA: IEEE, 2017: 319-328.
9 KIM T, KIM Y H, KIM I, et al. BASN: enriching feature representation using bipartite auxiliary supervisions for face anti-spoofing[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. Seoul, Korea: IEEE, 2019: 494-503.
10 GEORGE A, MARCEL S. Deep pixel-wise binary supervision for face presentation attack detection[C]// Proceedings of the 2019 International Conference on Biometrics (ICB). Crete, Greece: IEEE, 2019: 19352833.
11 RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[C]//Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention. Munich, Germany: Springer, 2015: 234-241.
12 SANDLER M, HOWARD A, ZHU M, et al. MobileNetV2: inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018: 4510-4520.
13 VISIN F, KASTNER K, CHO K, et al. ReNet: a recurrent neural network based alternative to convolutional networks[EB/OL]. (2015-07-23)[2023-02-21]. https://arxiv.org/abs/1505.00393.
14 BOULKENAFET Z, KOMULAINEN J, LI L, et al. OULU-NPU: a mobile face presentation attack database with real-world variations[C]//Proceedings of the 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017). Washington, USA: IEEE, 2017: 612-618.
15 CHINGOVSKA I, ANJOS A, MARCEL S. On the effectiveness of local binary patterns in face anti-spoofing[C]//2012 BIOSIG-Proceedings of the International Conference of Biometrics Special Interest Group (BIOSIG). Darmstadt, Germany: IEEE, 2012: 13029854.
16 ZHANG Z, YAN J, LIU S, et al. A face anti-spoofing database with diverse attacks[C]//Proceedings of the 2012 5th IAPR International Conference on Biometrics (ICB). New Delhi, India: IEEE, 2012: 26-31.
17 LIU Y, JOURABLOO A, LIU X. Learning deep models for face anti-spoofing: binary or auxiliary supervision[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018: 389-398.
18 BOULKENAFET Z, KOMULAINEN J, AKHTAR Z, et al. A competition on generalized software-based face presentation attack detection in mobile scenarios[C]//Proceedings of the 2017 IEEE International Joint Conference on Biometrics (IJCB). Denver, USA: IEEE, 2017: 688-696.
19 JOURABLOO A, LIU Y, LIU X. Face de-spoofing: anti-spoofing via noise modeling[C]//Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: IEEE, 2018: 290-306.
20 LIU Y, STEHOUWER J, LIU X. On disentangling spoof trace for generic face anti-spoofing[C]//Proceedings of the European Conference on Computer Vision. Glasgow, England: Springer, 2020: 406-422.
[1] 王旭晴,魏伟波,杨光宇,宋金涛,吕婷,潘振宽. 基于算法展开的图像盲去模糊深度学习网络[J]. 山东大学学报 (工学版), 2023, 53(6): 35-46.
[2] 王碧瑶,韩毅,崔航滨,刘毅超,任铭然,高维勇,陈姝廷,刘嘉巍,崔洋. 基于图像的道路语义分割检测方法[J]. 山东大学学报 (工学版), 2023, 53(5): 37-47.
[3] 周晓昕,廖祝华,刘毅志,赵肄江,方艺洁. 融合历史与当前交通流量的信号控制方法[J]. 山东大学学报 (工学版), 2023, 53(4): 48-55.
[4] 于畅,伍星,邓秋菊. 基于深度学习的多视角螺钉缺失智能检测算法[J]. 山东大学学报 (工学版), 2023, 53(4): 104-112.
[5] 宋佳芮,陈艳平,王凯,黄瑞章,秦永彬. 基于Affix-Attention的命名实体识别语义补充方法[J]. 山东大学学报 (工学版), 2023, 53(2): 70-76.
[6] 袁钺,王艳丽,刘勘. 基于空洞卷积块架构的命名实体识别模型[J]. 山东大学学报 (工学版), 2022, 52(6): 105-114.
[7] 李旭涛,杨寒玉,卢业飞,张玮. 基于深度学习的遥感图像道路分割[J]. 山东大学学报 (工学版), 2022, 52(6): 139-145.
[8] 孟令灿,聂秀山,张雪. 基于遮挡目标去除的公交车拥挤度分类算法[J]. 山东大学学报 (工学版), 2022, 52(4): 83-88.
[9] 杨霄,袭肖明,李维翠,杨璐. 基于层次化双重注意力网络的乳腺多模态图像分类[J]. 山东大学学报 (工学版), 2022, 52(3): 34-41.
[10] 王心哲,邓棋文,王际潮,范剑超. 深度语义分割MRF模型的海洋筏式养殖信息提取[J]. 山东大学学报 (工学版), 2022, 52(2): 89-98.
[11] 蒋桐雨, 陈帆, 和红杰. 基于非对称U型金字塔重建的轻量级人脸超分辨率网络[J]. 山东大学学报 (工学版), 2022, 52(1): 1-8.
[12] 吴建清,宋修广. 同步定位与建图技术发展综述[J]. 山东大学学报 (工学版), 2021, 51(5): 16-31.
[13] 柴庆发,孙守晶,邱吉福,陈明,魏振,丛伟. 气象灾害条件下电网应急物资预测方法[J]. 山东大学学报 (工学版), 2021, 51(3): 76-83.
[14] 杨修远,彭韬,杨亮,林鸿飞. 基于知识蒸馏的自适应多领域情感分析[J]. 山东大学学报 (工学版), 2021, 51(3): 15-21.
[15] 廖锦萍,莫毓昌,YAN Ke. 基于C-LSTM的短期用电预测模型和应用[J]. 山东大学学报 (工学版), 2021, 51(2): 90-97.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] 季涛,高旭,孙同景,薛永端,徐丙垠 . 铁路10 kV自闭/贯通线路故障行波特征分析[J]. 山东大学学报(工学版), 2006, 36(2): 111 -116 .
[2] 丑武胜 王朔. 大刚度环境下力反馈主手自适应算法研究[J]. 山东大学学报(工学版), 2010, 40(1): 1 -5 .
[3] 李国正1,史淼晶1,李福凤2,王忆勤2. 舌体图像分割技术的实验分析与改进[J]. 山东大学学报(工学版), 2010, 40(5): 87 -95 .
[4] 邓斌,王江 . 基于混沌同步与自适应控制的神经元模型参数估计[J]. 山东大学学报(工学版), 2007, 37(5): 19 -23 .
[5] 高厚磊 田佳 杜强 武志刚 刘淑敏. 能源开发新技术——分布式发电[J]. 山东大学学报(工学版), 2009, 39(5): 106 -110 .
[6] 张建明, 刘泉声, 唐志成, 占婷, 蒋亚龙. 考虑剪切变形历史影响的节理峰值剪切强度准则[J]. 山东大学学报(工学版), 0, (): 77 -81 .
[7] 贾超,赵建宇,徐帮树,岳长城,李树忱 . 清水隧道围岩软土振动液化研究[J]. 山东大学学报(工学版), 2008, 38(1): 83 -87 .
[8] 齐辉 商庆森 朱海波 崔新壮 刘超.
采空区公路沥青路面结构层应力关键影响因素分析
[J]. 山东大学学报(工学版), 2009, 39(6): 121 -124 .
[9] 贠汝安1,2,董增川1,王好芳2. 基于NSGA2的水库多目标优化[J]. 山东大学学报(工学版), 2010, 40(6): 124 -128 .
[10] 雷小锋1,庄伟1,程宇1,丁世飞1,谢昆青2. OPHCLUS:基于序关系保持的层次聚类算法[J]. 山东大学学报(工学版), 2010, 40(5): 48 -55 .