PC PIN [4] PIN PIN n 10n 3 2 PIN Fig. 2 Feature indices for PIN input on touch-screen a = (a 1, a 2,, a i,, a 10n 3 ) (a i, i {1,, 2n 1}) : 2

Size: px

Start display at page:

Download "PC PIN [4] PIN 2. 2.1 2 PIN n 10n 3 2 PIN Fig. 2 Feature indices for PIN input on touch-screen a = (a 1, a 2,, a i,, a 10n 3 ) (a i, i {1,, 2n 1}) : 2"

そうほがり
7 years ago
Views:

マルチメディア分散協調とモバイル (DICOMO2014)シンポジウム平成26年7月 PIN 入力タッチスクリーンバイオメトリクスにおける識別手法の影響泉将之1 西村友佑2 柏木まもる3 佐村敏治4 西村治彦5 概要本研究ではスマートフォンを用いた PIN 入力タッチスクリーンバイオメトリクスについて検討を行なった

1 マルチメディア分散協調とモバイル (DICOMO2014)シンポジウム平成26年7月 PIN 入力タッチスクリーンバイオメトリクスにおける識別手法の影響泉将之1 西村友佑2 柏木まもる3 佐村敏治4 西村治彦5 概要本研究ではスマートフォンを用いた PIN 入力タッチスクリーンバイオメトリクスについて検討を行なった従来のキーストローク認証では得られなかったセンサやタッチスクリーンからの情報を利用した新たな特徴量を導入することにより認証精度の向上を図った識別手法には統計的手法である Euclidean Distance ED 法 Manhattan Distance MD 法を機械学習手法である Support Vector Machine SVM Back Propagation Neural Networks BPNN Learning Vector Quantization LVQ を用いる 21 名の被験者を対象に PIN 入力データを収集解析を行い特徴量の組み合わせによる認証率及びそのプロファイル数依存性について実験を行ったその結果統計的手法では PIN4 桁において EER 6.7 % PIN10 桁において EER 3.3 % 機械学習手法では EER % PIN10 桁において EER %という結果を得た Influence of Identification Methods by Touch-screen Biometrics for PIN Input Izumi Masayuki1 Nishimura Yusuke2 Kashiwagi Mamoru3 Samura Toshiharu4 Nishimura Haruhiko5 1. はじめに現在スマートフォンの利用者は年々増加しておりそれに伴い不正利用や情報漏洩の危険性も増大しつつある従来のスマートフォンのセキュリティとして PIN Personal Identification Number 入力図 1 左やパターン入力図 1 右パスワード入力などによる認証システムがあげられるしかしこれらは総当たり攻撃やショルダーハッキング画面に残った皮脂をなぞるなどの方法で比較的容易になりすましが可能である一方で指紋認証を搭載したスマートフォン端末も登場しているが複製された指紋によ図1 りなりすましができることが報告されている [1] その対策の 1 つとして PIN 入力やパスワード入力に対スマートフォンで用いられる認証システム左 : PIN ロック右 : パターンロック Fig. 1 User authentication system on smartphone. (left): PIN lock, (right): pattern lock 明石工業高等専門学校専攻科 Advanced Course, Akashi National College of Technology 大阪大学基礎工学部 School of Engineering Science, Osaka University NTT スマートコネクト NTT SmartConnect Corporation 明石工業高等専門学校電気情報工学科 Department of Electrical and Computer Engineering, Akashi National College of Technology 兵庫県立大学応用情報科学研究科 Graduate School of Applied Informatics, University of Hyogo するキーストローク認証が注目されているキーストローク認証は主に PC のキーボードにおいて文字入力時のパターンを利用したバイオメトリクスである [2] PIN 入力認証時のキーストロークデータによる認証の研究は 1980 年頃から行われていた [3] PIN の照会とキーストローク認証を組み合わせることでより強固な認証システムを構築 1044

2 PC PIN [4] PIN PIN n 10n 3 2 PIN Fig. 2 Feature indices for PIN input on touch-screen a = (a 1, a 2,, a i,, a 10n 3 ) (a i, i {1,, 2n 1}) : 2n 1 (a i, i {2n,, 4n 2}) : 2n 1 (a i, i {4n 1,, 5n 3}) : n 1 x y (a i, i {5n 2,, 9n 3}) x y : 4n (a i, i {9n 2,, 10n 3}) : n Registration: 3 Authentication: 3 3 Fig. 3 Authentication process of Touch-screen biometrics PIN 4 m j a j = a i j, i {1, 2,, 10n 3}, j {1, 2,, m} a i σ i a i = 1 m a ik (1) m k=1 1 m σ i = (a ik a i ) m 2 (2) k=1 (3)

3 0105,0,0, , , , ,0,1, , , , ,1,0, , , , ,1,1, , , , ,0,0, , , , ,0,1, , , , ,5,0, , , , ,5,1, , , , ,OK,0, , , , ,OK,1,68.476, , , Fig. 4 PIN Example of collected PIN input data 5 Fig. 5 ED : : Authentication by ED method (left): correct user, (right): incorrect user a i (4) a i j = a i j a i (3) σ i a i = 1 m a ik (4) m u i (3) (5) k=1 d ED = 10n 3 1 (u i 10n 3 a i )2 (6) d ED T H i=0 d ED dt ED H (7) d ED > dt ED H MD MD 6 u i = u i a i σ i (5) ED: Euclidean Distance MD: Manhattan Distance SVM: Support Vector Machine BPNN: Back Propagation Neural Networks LVQ: Learning Vector Quantization (5) u i ED ED 5 ED d ED (6) 6 MD : : Fig. 6 Authentication by MD method (left): correct user, (right): incorrect user MD d MD (8) d MD = dt MD H 10n 3 1 u i 10n 3 a i (8) i=0 ED d MD dt MD H (9) d MD > dt MD H

4 u i SVM SVM: Support Vector Machine [5], [6] t ωx i + ω i = 1 t ωx j + ω i = Fig. 8 Maximum margin hyperplane and margins with samples from two classes x i x j = 2/ ω ω (12) ω 2 2 y i ( t ωx i + ω 0 ) 1 (12) Fig. 7 7 Samples that are linearly separable by hyperplane of multiple (x 1, y 1 ), (x 2, y 2 ),, (x n, y n ) x i (X = {x 1, x 2,, x n }) y i 1 y i = 1 2 y i = 1 y i (10) 1 ( t ωx i + ω 0 0) y i = 1 ( t ωx i + ω 0 < 0) (10) x i ω d ω ω 0 2 (11) 8 SVM (13) y i ( t ωx i + ω 0 ) 1 ξ i (ξ i 0) (13) ξ i (13) (14) C t ωx + b = 0 (11) (11) 2 ω ω 0 SVM ω 2 N + C ξ i C > 0 (14) 2 i=1 SVM

5 ϕ(x i ) K(x i, x j ) t x i x j (15) (γ t x i x j + δ) p (16) exp( γ t x i x j ) (17) tanh(γ t x i x j δ) (18) γ p SVM SVM BPNN NN: Neural Networks [5], [7] 9 x 1, x 2,, x n 9 Fig. 9 Neuron model w 1, w 2,, w n x i w i u = w i x i f (u) (19) 1 n y = f (u) = 1 + e u u = w i x i (19) i=1 9 w i Fig. 10 Hierarchical neural network of three layers 1 BPNN: Back Propagation Neural Networks BPNN w i x i BPNN LVQ LVQ: Learning Vector Quantization SOM: Self- Organizing Maps [8] LVQ1 LVQ2 LVQ3 LVQ LVQ

2.3.2.3.1 LVQ1 LVQ1 (x, y) c y l c (21) c = arg min x m i (20) c(t) + α(x c(t)) y = l c c(t + 1) = (21) c(t) α(x c(t)) y l c 2.3.2.3.2 LVQ2 LVQ2 LVQ1 2 c 1 c 2 l i l j l j = y c i (t + 1) = c i (t) α(x c i (t)) (29) c j (t + 1) = c j (t) + α(x c j (t)) (30) LVQ 3.

6 LVQ1 LVQ1 (x, y) c y l c (21) c = arg min x m i (20) c(t) + α(x c(t)) y = l c c(t + 1) = (21) c(t) α(x c(t)) y l c LVQ2 LVQ2 LVQ1 2 c 1 c 2 l i l j l j = y c i (t + 1) = c i (t) α(x c i (t)) (29) c j (t + 1) = c j (t) + α(x c j (t)) (30) LVQ PIN 3 PIN Android 11 c = arg min x m i (22) c = arg min x m j (23) c 1 c % (24) (25) c 1 (t + 1) = c 1 (t) + α(x c 1 (t)) (24) c 2 (t + 1) = c 2 (t) α(x c 2 (t)) (25) LVQ3 LVQ3 LVQ1 LVQ2 2 c i = arg min x m i (26) c j = arg min x m j (27) 2 l i l j y l i = l j = y c i, j (t + 1) = c i, j (t) + α(x c i, j (t)) (28) 11 PIN : PIN : Fig. 11 Screenshot from interface of pin input data collecting system. (left): PIN input screen, (right): data transmission confirmation screen PIN 10 PIN PIN 48 FRR: False Rejection RateFAR: False Acceptance Rate FRR FAR EER: Equal Error Rate

7 EER EER FRR FAR R SVM e1071 svm tune.svm C = 10 i (i 0, 0.2,, 2.0) γ = 10 j ( j 3.0, 2.8,, 1.0) BPNN nnet nnet 5 LVQ nnet lvq3 LVQ3 pixel dp Density-independent Pixel dp 160dpi dots per inch [9] dp = pixel (160/dpi) (31) PIN A 1 10 PIN A 2 1 EER 1 PIN4 PIN10 2 FRR FAR 4 PIN MD dt MD H PIN MD dt MD H 13 PIN d MD T H = FRR FAR EER Error Rate [%] Fig. 12 Error Rate [%] Fig d MD TH FRR FAR PIN4 MD Dependence on threshold of recognition accuracy (4-digits PIN, MD method d MD TH FRR FAR PIN10 MD Dependence on threshold of recognition accuracy (10-digits PIN, MD method PIN SVM A 3 BPNN A 4 LVQ A 5 10 PIN SVM A 6 BPNN A 6 LVQ A

8 PIN PIN MD PIN MD 15 EER EER [%] # of profiles 14 EER PIN4 MD Fig. 14 Dependence on the number of profiles in the EER curve of recognition accuracy(4-digit PIN, MD method) EER [%] # of profiles 15 EER PIN10 MD Fig. 15 Dependence on the number of profiles in the EER curve of recognition accuracy(10-digit PIN, MD method) SVM BPNN LVQ FRR FAR 4 PIN SVM A 9 BPNN A 10 LVQ A PIN SVM A 12 BPNN A 13 LVQ A 14 PIN FRR FAR FRR FAR PDA PIN 1 Clarke Furnell[10] 4 11 PIN 6 FF-MLP Feed Forword Multi-Layer Perceptron RBF Radial Basis Function Networks Generalized Regression Neural Network GRNN Saevanee Bhatarakosol[11] PDA Personal Digital Assistant k k-nn: k-nearest neighbors algorithm Hwang [12] 4. PIN

9 1 Table 1 Comparison with related research PIN PIN EER [%] Clarke and Furnell[10] Saevanee and Bhatarakosol[11] k 1 Hwang et al.[12] ED MD SVM BPNN LVQ 7 PIN [8] T. (2005). [9] : Supporting Multiple Screens Android Developers, Google Inc. (online), available from support.html (accessed ). [10] Clarke, N. L. and Furnell, S.: Advanced user authentication for mobile devices, computers & security, Vol. 26, No. 2, pp (2007). [11] Saevanee, H. and Bhatarakosol, P.: User authentication using combination of behavioral biometrics over the touchpad acting like touch screen of mobile device, Computer and Electrical Engineering, ICCEE International Conference on, IEEE, pp (2008). [12] Hwang, S.-s., Cho, S. and Park, S.: Keystroke dynamicsbased authentication for mobile devices, Computers & Security, Vol. 28, No. 1, pp (2009). [1] : CCC Chaos Computer Club breaks Apple TouchID, Chaos Computer Club (online), available from (accessed ). [2] Samura, T. and Nishimura, H.: Personal Identification and Authentication Based on Keystroke Dynamics in Japanese Long-Text, Continuous Authentication based on Biometrics: Data, Models, and Metrics, I. Traore et al.(eds.), IGI Global, pp (2011). [3] Banerjee, S. P. and Woodard, D. L.: Biometric authentication and identification using keystroke dynamics: A survey, Journal of Pattern Recognition Research, Vol. 7, No. 1, pp (2012). [4] 3 pp (2013). [5] : (2007). [6] C. M. : (2008). [7] C. M. : (2008).

10 A.1 A 1 PIN4 EER Table A 1 EER for combination of features by statistics method in 4 digits PIN ED EER [%] MD EER [%] A 2 PIN10 EER Table A 2 Recognition accuracy for combination of features by statistics method in 10 digits PIN ED EER [%] MD EER [%]

11 A 3 PIN4 SVM Table A 3 Recognition accuracy for combination of features by SVM in 4 digits PIN SVM FRR [%] SVM FAR [%] A 4 PIN4 BPNN Table A 4 Recognition accuracy for combination of features by BPNN in 4 digits PIN BPNN FRR [%] BPNN FAR [%]

12 A 5 PIN4 LVQ Table A 5 Recognition accuracy for combination of features by LVQ in 4 digits PIN LVQ FRR [%] LVQ FAR [%] A 6 PIN10 SVM Table A 6 Recognition accuracy for combination of features by SVM in 10 digits PIN SVM FRR [%] SVM FAR [%]

13 A 7 PIN10 BPNN Table A 7 Recognition accuracy for combination of features by BPNN in 10 digits PIN BPNN FRR [%] BPNN FAR [%] A 8 PIN10 LVQ Table A 8 Recognition accuracy for combination of features by LVQ in 10 digits PIN LVQ FRR [%] LVQ FAR [%]

14 A.2 Table A 9 A 9 SVM PIN4 Dependence on the number of training data of recognition accuracy using SVM (4-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%] Table A 10 A 10 BPNN PIN4 Dependence on the number of training data of recognition accuracy using BPNN (4-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%] Table A 11 A 11 LVQ PIN4 Dependence on the number of training data of recognition accuracy using LVQ (4-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%]

15 Table A 12 A 12 SVM PIN10 Dependence on the number of training data of recognition accuracy using SVM (10-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%] Table A 13 A 13 BPNN PIN10 Dependence on the number of training data of recognition accuracy using BPNN (10-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%] Table A 14 A 14 LVQ PIN10 Dependence on the number of training data of recognition accuracy using LVQ (10-digits PIN) FRR [%] FAR [%] FRR [%] FAR [%] FRR [%] FAR [%]

一般 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS Proceedings of Biometrics Workshop

一般 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS Proceedings of Biometrics Workshop 679-3 679-3 679-3 E-mail: samura@akashi.ac.jp, masayuki@izumin.info, haru@ai.u-hyogo.ac.jp 4