Face anti-spoofing (FAS) is significant for the security of face recognition systems. neural networks (NNs), including convolutional neural network (CNN) and vision transformer (ViT), have been dominating the field of the FAS. However, NN-based methods are vulnerable to adversarial attacks. Attackers could insert adversarial noise into spoofing examples to circumvent an NN-based face-liveness detector. Our experiments show that the CNN or ViT models could have at least an 8% equal error rate (EER) increment when encountering adversarial examples. Thus, developing methods other than NNs is worth exploring to improve security at the system level. In this paper, we have proposed a novel solution for FAS against adversarial attacks, leveraging a deep forest model. Our approach introduces a multi-scale texture representation based on local binary patterns (LBP) as the model input, replacing the grained-scanning mechanism (GSM) used in the traditional deep forest model. Unlike GSM, which scans raw pixels and lacks discriminative power, our LBP-based scheme is specifically designed to capture texture features relevant to spoofing detection. Additionally, transforming the input from the RGB space to the LBP space enhances robustness against adversarial noise. Our method achieved competitive results. When testing with adversarial examples, the increment of EER was less than 3%, more robust than CNN and ViT. On the benchmark database IDIAP REPLAY-ATTACK, a 0% EER was achieved. This work provides a competitive option in a fusing scheme for improving system-level security and offers important ideas to those who want to explore methods besides CNNs. To the best of our knowledge, this is the first attempt at exploiting the deep forest model in the problem of FAS, with the consideration of adversarial attacks.
Citation: Rizhao Cai, Liepiao Zhang, Changsheng Chen, Yongjian Hu, Alex Kot. Learning deep forest for face anti-spoofing: An alternative to the neural network against adversarial attacks[J]. Electronic Research Archive, 2024, 32(10): 5592-5614. doi: 10.3934/era.2024259
Related Papers:
[1]
Samuel Asante Gyamerah, Bright Emmanuel Owusu, Ellis Kofi Akwaa-Sekyi .
Modelling the mean and volatility spillover between green bond market and renewable energy stock market. Green Finance, 2022, 4(3): 310-328.
doi: 10.3934/GF.2022015
[2]
Aikaterini (Katerina) Tsoukala, Georgios Tsiotas .
Assessing green bond risk: an empirical investigation. Green Finance, 2021, 3(2): 222-252.
doi: 10.3934/GF.2021012
[3]
Vilija Aleknevičienė, Raimonda Vilutytė .
Short-term stock market reaction to the announcement of green bond issue: evidence from Nordic countries. Green Finance, 2024, 6(4): 728-744.
doi: 10.3934/GF.2024028
[4]
Samuel Asante Gyamerah, Clement Asare .
A critical review of the impact of uncertainties on green bonds. Green Finance, 2024, 6(1): 78-91.
doi: 10.3934/GF.2024004
[5]
Tao Lin, Mingyue Du, Siyu Ren .
How do green bonds affect green technology innovation? Firm evidence from China. Green Finance, 2022, 4(4): 492-511.
doi: 10.3934/GF.2022024
[6]
Toyo Kawabata .
Private governance schemes for green bond standard: influence on public authorities’ policy making. Green Finance, 2020, 2(1): 35-54.
doi: 10.3934/GF.2020003
[7]
Vitor Miguel Ribeiro .
Green bond market boom: did environmental, social and governance criteria play a role in reducing health-related uncertainty?. Green Finance, 2023, 5(1): 18-67.
doi: 10.3934/GF.2023002
[8]
Kentaka Aruga, Md. Monirul Islam .
Assessing investor preferences for environmental and sustainability bonds in Japan: A discrete choice experiment approach. Green Finance, 2025, 7(2): 200-222.
doi: 10.3934/GF.2025008
[9]
Shahinur Rahman, Iqbal Hossain Moral, Mehedi Hassan, Gazi Shakhawat Hossain, Rumana Perveen .
A systematic review of green finance in the banking industry: perspectives from a developing country. Green Finance, 2022, 4(3): 347-363.
doi: 10.3934/GF.2022017
[10]
Fernando Palao, Ángel Pardo .
Carbon and safe-haven flows. Green Finance, 2022, 4(4): 474-491.
doi: 10.3934/GF.2022023
Abstract
Face anti-spoofing (FAS) is significant for the security of face recognition systems. neural networks (NNs), including convolutional neural network (CNN) and vision transformer (ViT), have been dominating the field of the FAS. However, NN-based methods are vulnerable to adversarial attacks. Attackers could insert adversarial noise into spoofing examples to circumvent an NN-based face-liveness detector. Our experiments show that the CNN or ViT models could have at least an 8% equal error rate (EER) increment when encountering adversarial examples. Thus, developing methods other than NNs is worth exploring to improve security at the system level. In this paper, we have proposed a novel solution for FAS against adversarial attacks, leveraging a deep forest model. Our approach introduces a multi-scale texture representation based on local binary patterns (LBP) as the model input, replacing the grained-scanning mechanism (GSM) used in the traditional deep forest model. Unlike GSM, which scans raw pixels and lacks discriminative power, our LBP-based scheme is specifically designed to capture texture features relevant to spoofing detection. Additionally, transforming the input from the RGB space to the LBP space enhances robustness against adversarial noise. Our method achieved competitive results. When testing with adversarial examples, the increment of EER was less than 3%, more robust than CNN and ViT. On the benchmark database IDIAP REPLAY-ATTACK, a 0% EER was achieved. This work provides a competitive option in a fusing scheme for improving system-level security and offers important ideas to those who want to explore methods besides CNNs. To the best of our knowledge, this is the first attempt at exploiting the deep forest model in the problem of FAS, with the consideration of adversarial attacks.
1.
Introduction
Entropy is a fundamental uncertainty measure of a random variable. Shannon [1] provided a quantitative measure of uncertainty as a measure of information. A flexible extension of Shannon entropy was introduced by Rényi [2]. The Rényi entropy is one parameter generalization of the Shannon entropy and can be used as a measure of randomness. Applications of Rényi entropy can be found in many fields, such as biology, genetics, electrical engineering, computer science, economics, chemistry and physics.
The Rényi entropy has been widely used in analysis of the quantum systems. It can be employed in the analysis of quantum communication protocols (Renner et al. [3]), quantum correlations (Lévay et al. [4]). Also it can be used to characterize the number of reads needed to reconstruct a DNA sequences (see, Motahari et al. [5]). Rényi entropy has numerous issues as a data analysis tools in many practical applications, especially in data that dealing with time-frequency representations. Gabarda and Cristobal [6] applied the Rényi entropy to an image fusion method and identified which pixels have a higher amount of information among the given input images. Rényi entropy can be used to calculate the risk measure of portfolio.
Entropy concepts can be related to informational market efficiency. Martina et al. [7] studied the complexity of crude oil prices using entropy measures to monitor the evolution of crude oil price movements. They found that the methods based on entropy concepts can shed light on the structure of crude oil markets as well as on its link to macroeconomic conditions and socio-political extreme events. Resconi et al. [8] provided a new suggestive reading of quantum mechanics by starting from the superposition of different Boltzmann entropies. They suggested that a quantum entropy space can be used as the fundamental arena that describing the quantum effects. Another measure of uncertainty deals with residual lifetime function is dynamic cumulative residual entropy (DCRE) which can be attractive in many fields like reliability and survival analysis. Recently, there are considerable literatures assigned to the applications, generalizations and properties of Rényi's measure of entropy.
Recently, there are many literatures studied the estimation procedure for the entropy measures. Kayal and Kumar [9] considered Bayesian estimation of entropy for exponential distribution under LINEX loss function. Seo et al. [10] produced an entropy estimate using upper record values from the generalized half-logistic distribution. Based on generalized progressive hybrid censoring scheme, Bayesian estimates of entropy for Weibull distribution were discussed by Cho et al. [11]. Based on record values, Chacko and Asha [12] studied the estimation of the entropy for generalized exponential distribution. Patra et al. [13] estimated a function of scale parameter of an exponential population under general loss function. Maximum likelihood (ML) estimate of Shannon entropy for inverse Weibull distribution under multiple censored data was discussed by Hassan and Zaky [14]. Petropoulos et al. [15] provided improved estimators of the entropy for the scale mixture of exponential distributions. Estimation of entropy for inverse Lomax distribution under multiple censored has been studied by Bantan et al. [16]. Bayesian estimate of the entropy for Lomax distribution based on upper record values has been studied by Hassan and Zaky [17].
If X is an absolutely continuous random variable with probability density function (PDF) f (x), then the corresponding Rényi entropy of order β is defined as:
IR(β)=11−βlog(∞∫−∞fβ(x)dx),β>0andβ≠1.
(1)
Recently, measurements of uncertainty for probability distributions became more interested. Sunoj and Linu [18] defined the cumulative residual Rényi entropy (CRRE) for residual lifetime Xt = (X −t | X > t) based on survival function instead of using PDF as follows:
γR(β)=11−βlog(∞∫0ˉFβ(x)dx),β>0andβ≠1,
(2)
where, ˉF(t)=P(X>t)=1−F(t) is the survival (reliability) function. The main features of the CRRE are always non-negative as well as it has consistent definitions in both the continuous and discrete domains. Furthermore it can be easily computed from sample data and these computations asymptotically converge to the true values (see, Rao et al. [19]). Sunoj and Linu [18] defined the DCRRE as a dynamic form of uncertainty as follows:
γR(β)=11−βlog(∞∫tˉFβ(x)ˉFβ(t)dx),β>0andβ≠1.
(3)
Therefore, when t = 0, the DCRRE tends to CRRE. In the literature, few studies had been done concerning the inferential procedures of the entropy measures incorporating DCRRE for lifetime distributions. Kamari [20] produced some properties of the DCRRE based on order statistics. Kundu et al. [21] presented the cumulative residual and past inaccuracy measures which are extensions of the corresponding cumulative entropies for truncated random variables. Renjini et al. [22] discussed the Bayesian estimates of the DCRRE for the Pareto distribution using Type II right censored data. Renjini et al. [23] discussed the Bayesian estimates of the DCRRE for Pareto distribution of the first kind under upper record values. Renjini et al. [24] obtained Bayesian estimate of the DCRRE for Pareto distribution based on complete data. Bayesian estimate of dynamic cumulative residual Shannon entropy for Pareto II distribution has been studied by Ahmadini et al. [25].
Lomax [26] proposed a very good alternative to the common lifetime distributions such as exponential, Weibull, or gamma when the experimenter presumes that the population distribution may be heavy-tailed. Lomax distribution can be applied in a variety of fields such as a business failure data model, income and wealth inequality, computer science, risk analysis and economics, actuarial science and reliability. Corbellini et al. [27] used it to model firm size and queuing problems. As an important lifetime distribution, the Lomax distribution with shape parameter α and scale parameter λ has the following PDF
f(x;α,λ)=αλα(x+λ)−(α+1),x,α,λ>0.
(4)
The cumulative distribution function (CDF) corresponding to (4) is given by,
F(x;α,λ)=1−λα(x+λ)−α,x,α,λ>0.
(5)
At λ=1,α≠1, Lomax distribution reduces to the beta prime distribution (also known as inverted beta distribution or beta distribution of the second kind). At α=1,λ≠1, Lomax distribution reduces to a log-logistic distribution. At α=λ=1, the Lomax distribution reduces to F-distribution, F (2, 2). The reliability and hazard rate functions of Lomax distribution are given, respectively, by
ˉF(x;α,λ)=λα(x+λ)−α,
and,
h(x;α,λ)=α(x+λ)−1.
The hazard rate function of Lomax distribution takes different shapes according to values of shape parameter α. It is a constant at α=1, and it is continuously decreasing at α>1, which represents early failures.
Studies about the Lomax distribution have been provided by many authors. For instance, Abd-Elfattah et al. [28] obtained the Bayesian and non-Bayesian estimates of the sample size for the Lomax distribution in case of Type-I censored samples. Record values of Lomax distribution were discussed by Ahsanullah [29]. Balakrishnan and Ahsanullah [30] introduced some recurrence relations between the moments of record values from Lomax distribution. Hassan and Al-Ghamdi [31] determined the optimum test plan for simple step stress accelerated life testing. Hassan et al. [32] discussed the optimal times of changing stress level for k-level step stress accelerated life tests based on adaptive Type-II progressive hybrid censoring with product's life time following Lomax distribution. For more application about Lomax distribution, the reader can refer to [33,34,35,36,37,38].
From the above, it is evident that the Lomax distribution has been received greatest attention from theoretical and statisticians primarily due to its use in various fields. In the literature, there are no reports about the statistical inference of the DCRRE for Lomax distribution. So we need to estimate the DCRRE for Lomax distribution in view of Bayesian and non-Bayesian procedures based on complete samples. The ML estimates and approximate confidence intervals of the DCRRE are obtained. The Bayesian estimate is calculated using gamma priors under squared error (SE), LINEX and precautionary (PRE) loss functions. According to the complicated forms of the DCRRE Bayesian estimate, we employ the Markov Chain Monte Carlo (MCMC) simulation technique for numerical study.
This article can be organized as follows. Section 2 gives the DCRRE of Lomax distribution as well as obtains the ML and approximate confidence intervals of DCRRE. Section 3 presents Bayesian estimate of the DCRRE for Lomax distribution under symmetric and asymmetric loss functions. The MCMC technique is given in Section 4. A real data application is illustrated in Section 5. The paper ends with some conclusions based on the results of the numerical studies.
2.
Maximum likelihood estimate of the DCRRE
This section provides an explicit expression of the DCRRE for Lomax distribution. Then, the ML estimate of the DCRRE is derived based on complete random sample.
The DCRRE of Lomax distribution is obtained by substituting (5) in (3) as follows:
γR(β)=11−βlog(∞∫t(x+λ)−αβ(t+λ)−αβdx).
(6)
Hence, after simplification the DCRRE takes the form
γR(β)=11−βlog(t+λαβ−1),β>0,β≠1andαβ>1.
(7)
This is the required expression of the DCRRE for Lomax distribution. To obtain the ML estimate of γR(β), we must obtain the ML estimate of the model parameters α and λ.
Consider a simple random sample of size n drawn from PDF (4) and CDF (5), where α and λ are unknown. Then, given the sample x_=(x1,x2,...,xn), the likelihood function of Lomax distribution can be written as follows
L(α,λ|x_)=αnλnαn∏i=1(xi+λ)−(α+1).
(8)
It is usually easier to maximize the natural logarithm of the likelihood function rather than the likelihood function itself. Therefore, the logarithm of likelihood function, say lnℓ, is
lnℓ=nlogα+nαlogλ−(α+1)n∑i=1log(xi+λ).
The partial derivatives of the log-likelihood function with respect to α and λ, are obtained as follows:
∂lnℓ∂α=nα+nlogλ−n∑i=1log(xi+λ),
(9)
and
∂lnℓ∂λ=nαλ−(α+1)n∑i=1(xi+λ)−1.
(10)
The ML estimates of α and λ are determined by solving the equations ∂lnℓ/∂lnℓ∂α∂α=0, and ∂lnℓ/∂lnℓ∂λ∂λ=0, simultaneously. Further the resulting equations cannot be solved analytically, so, numerical technique must be applied to solve these equations, simultaneously, to obtain ˆαML and ˆλML. The existence and uniqueness of ˆαML and ˆλML can be checked by solving, algebraically, ∂lnℓ/∂lnℓ∂α∂α=0 and ∂lnℓ/∂lnℓ∂λ∂λ=0, simultaneously and checking that, the second derivative is negative which is a maximum point.
Hence, by using the invariance property of ML estimate, then ML estimates of the DCRRE denoted by ˆγR,ML(β) becomes
Furthermore, a confidence interval of the DCRRE is the probability that a real value of the entropy will fall between an upper and lower bounds of a probability distribution. For large sample size, the ML estimates, under appropriate regularity conditions, are consistent and asymptotically normally distributed. Therefore, the two-sided approximate confidence limits for ˆγR,ML(β) can be constructed, such that
where, zη/η22 is 100 (1−η) the standard normal percentile, and η is the significant level. Therefore, the lower (L) and upper (U) confidence limits for ˆγR,ML(β) can be obtained as
where σ(ˆγR,ML(β)) is the standard deviation. The two-sided approximate confidence limits for DCRRE will be constructed with confidence level 95 %.
3.
Bayesian estimate of the DCRRE
In this section, the Bayesian estimates of γR(β) under symmetric (SE) and asymmetric (LINEX and PRE) loss functions are obtained by assuming the priors of parameters α and λ have gamma distributions. As reported by Kayal and Kumar [9], the SE loss function is used when weights of error caused by observation and underestimation are equal. But asymmetric loss function is used when the overestimation may have more serious than underestimation or vice versa. To compute the Bayesian estimate of γR(β), we proceed as follows. Firstly, the Bayesian estimates of α and λ are obtained and then the DCRRE estimate will be calculated by substituting the parameter estimates in (7). Additionally, the width of the Bayesian credible interval (BCI) estimates is obtained.
In Bayesian method, there are many methods that can be used in selection of prior. There are two types of the prior distributions, non-informative and informative prior (IP) distributions. The prior of parameters may be selected as independent or dependent priors. In this paper, we assume the prior of parameters α and λ, denoted by π1(α) and π2(λ) are independent. Following Pak and Mahmoudi [39], we take π1(α) and π2(λ) to be gamma distributions with parameters (a, b) and (c, d) as follows
π1(α)=abΓ(b)αb−1e−αa,π2(λ)=cdΓ(c2)λc−1e−λd,
(14)
where a, b, c and d are known and non-negative. So, the joint posterior for parameters, denoted by π∗1,2(α,λ), is
where, ν is a real number. Additionally, the Bayesian estimates of parameters α and λ under PRE loss function say ˆαPRE and ˆλPRE are given as follows:
Integrals (18)–(23) do not take a closed form, so the Metropolis-Hastings (M-H) and random-walk Metropolis algorithms are employed to generate the MCMC samples from posterior density functions (16) and (17), respectively. Hence, the Bayesian estimates of α and λ under SE, LINEX and PRE loss functions are obtained as the mean of the simulated samples from their posteriors. Further, once the Bayes estimates of α and λ are obtained. Therefore, the Bayesian estimate of the DCRRE, denoted by ˆγR,BE(β)SE, under SE loss function is obtained, by using (7), as follows:
ˆγR,BE(β)SE=11−βlog(t+ˆλSEˆαSEβ−1).
(24)
By similar way, we obtain the Bayesian estimate of γR(β) for Lomax distribution under LINEX, and PRE loss functions. Furthermore, the BCI is a useful summary of the posterior distribution which reflects its variation that is used to quantify the statistical uncertainty. An approximate highest posterior density interval of γR(β) is obtained by using the same algorithm of Chen and Shao [40].
4.
Monte Carlo simulation
This section evaluates the performance of the ML and Bayes estimates (BEs) based on simulation study.
The maximum likelihood estimates (MLEs) of the DCRRE for various sample sizes are computed by using Equations (9) and (10). Criteria measures include mean square error (MSE), average length (AL) and coverage probability (CP) are computed to investigate the performance of MLEs. Simulation procedures can be described as follows:
● 1000 random samples of sizes n = 30, 50, 70 and 100 are generated from Lomax distribution.
● The values of parameters are selected as α=(1.5,2.5),λ=(0.5,1.5,4) and β=(3,5).
● The true values of the DCRRE measure are selected as γR(β)= 0.2798, 0.6264 and 0.9359 at t = 0.5, while γR(β)= 0.077075, 0.2798 and 0.58932 at t = 1.5.
● The MLEs of α and λ are obtained from (9) and (10), then the MLE of γR(β) is obtained by substituting ˆαML and ˆλML in (11). The approximate confidence interval of ˆγR,ML(β) using (12) are constructed with confidence level at η=0.05.
● Compute the average MLEs of the DCRRE, MSEs, CPs, and ALs.
The M-H algorithm is one of the most famous subclasses of the MCMC method in Bayesian literature used to simulate the deviates from the posterior density and produce the good approximate results. The relative absolute biases (RABs), estimated risks (ERs) and the width of the BCI are computed to assess the behavior of the Bayesian estimates. A simulation study is done via R 3.1.2. Additionally, tables of simulation results are given. To compare the DCRRE Bayes estimates, the MCMC simulations are designed for different sample sizes under SE, LINEX and PRE loss functions. The values of parameters are selected as α=(1.5,2.5),λ=(0.5,1.5,4) and β=(3,5). The true values of the DCRRE measure are selected as γR(β)= 0.2798, 0.6264 and 0.9359 at t = 0.5, while γR(β)= 0.077075, 0.2798 and 0.58932 at t = 1.5. The hyper-parameters for gamma distribution are selected as a = c = 1 and b = d = 4. Also, we take (v = −2, 2) for LINEX loss function. Take N = 5000 random samples of sizes n, where each sample size (n) = 10, 30, 50, 70 and 100 are generated from Lomax distribution.
The M-H algorithm is described as follows:
Step 1: Let g(.) be the PDF of subject distribution;
Step 2: Initialize a starting value x0 and determine the number of samples N;
Step 3: for i = 2 to N set x=xi−1;
Step 4: Generate u from uniform (0, 1) and generate y from g(.);
Step 5: if u⩽πα(y)g(x)πα(x)g(y) then set xi=y else set xi=x;
Step 6: Set i=i+1 and return to step 2 and repeat the previous steps N times.
4.1. Numerical results based on ML method
Numerical observations for DCRRE estimates using ML method at β=3 and β=5 are presented in Tables 1–4 and represented in Figures 1–4. Some observations related to the behavior of the DCRRE estimates are outlined as follows:
Table 1.
MSE, AL and CP of 95% CI for DCRRE estimates of Lomax distribution at t = 0.5 and β=3.
● The value of MLE, ˆγR,ML(β), decreases as the value of λ increases for fixed value of α (see Tables 1 and 2)
● The value of MLE, ˆγR,ML(β), decreases as α decreases for fixed value of λ (see Tables 1 and 2).
● The value of MLE, ˆγR,ML(β), at α=1.5,λ=4 has the smallest value compared with the corresponding estimates of the other sets of parameters (see Tables 3 and 4).
● The MSE of ˆγR,ML(β) decreases as the sample size increases.
● For fixed value of λ, the MSE of ˆγR,ML(β) decreases as the value of α decreases.
● For fixed value of α, the MSE of ˆγR,ML(β) decreases as the value of λ increases.
● The MSE of the DCRRE estimates gets the smallest value at γR(β)=0.2798, where (α=1.5,λ=1.5),t = 0.5 and β=3, compared to the MSE of MLE of DCRRE measure for the corresponding other sets of parameters (see Figure 1).
● The MSE of the DCRRE estimates takes the smallest value at γR(β)=0.077075, where (α=1.5,λ=1.5),t = 1.5 and β=3, compared to MSE of MLE of DCRRE for the corresponding other sets of parameters (see Figure 2).
● At β=5,(α=1.5,λ=1.5), where the true value of γR(β)=0.2946637 the MSE of ˆγR,ML(β) has the smallest value compared to the MSE of MLEs for the corresponding other sets of parameters at t = 0.5 (see Figure 3).
● At t = 1.5, the MSE of ˆγR,ML(β) at (α=1.5,λ=4.0), has the smallest value compared to the MSE of MLEs for the corresponding other sets of parameters (see Figure 4).
Generally as seen from Tables 1–4 that the CP is very close to their corresponding nominal levels in approximately most of the cases. Also, the CP increases as the sample size increases for all values of DCRRE.
● The MLE of the DCRRE decreases as the value of λ increases.
● Regarding the AL of estimates, it can be observed that, as n increases the AL of DCRRE estimates decreases.
4.2. Numerical results based on Bayesian method
Tables 5–10 and Figures 5–8, list the numerical outcomes for different estimates of the DCRRE under different loss functions at β=3.Tables 11–14 and Figures 9–12, list and describe the simulation outcomes for different estimates of the DCRRE under different loss functions at β=5. Some observations related to the behavior of the DCRRE estimates are provided as follows:
Table 5.
Bayes estimates, RAB, ER and width of the DCRRE for Lomax distribution for (α,λ)=(1.5,1.5),t = 0.5 and β=3.
● As anticipated, the performance of all DCRRE estimates become improved as the sample size increases.
● The estimated value of the DCRRE increases as the value of α increases for fixed λ. The estimated value of the DCRRE decreases as the value of scale parameter λ increases for fixed α.
● The estimated value of the DCRRE decreases as the value of t increases for fixed values of α and λ.
● Under SE loss function, the ERs for DCRRE estimates get the smallest values at true value γR(β)= 0.6264, for most values of n (see for example Figures 5 and 6).
● The ER of ˆγR,BE(β)LINEX at v = −2 gets the smallest values for all values of n. Also, the width of the BCI for ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for all n (see Table 5).
● The ER of ˆγR,BE(β)SE takes the smallest values for all n except n = 100. The width of the BCI for ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of the PRE and SE loss functions for all n (see Table 6).
● The ER of ˆγR,BE(β)LINEX at v = −2 takes the smallest values for all values of n. The width of the BCI for ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of PRE and SE loss functions at n = 10 and 30 (see Table 7).
● At t = 1.5, the ERs for DCRRE estimates get the smallest values at γR(β)= 0.2798, for all n under SE and PRE loss functions (see for example Figures 7 and 8).
● The ER of ˆγR,BE(β)LINEX at v = 2 takes the smallest values at n = 50, 70 and 100, while the ER of ˆγR,BE(β)PRE takes the smallest values at n = 10 and 30. The width of the BCI for ˆγR,BE(β)SE is the shortest compared to the width of the BCI in case of LINEX and PRE loss functions for all values of n (see Table 8).
● The ER of ˆγR,BE(β)LINEX at v = −2 takes the smallest values at n = 50, 70 and 100, while the ER of ˆγR,BE(β)PRE takes the smallest values at n = 10 and 30. The width of the BCI of ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for all values of n (see Table 9).
● The ER of ˆγR,BE(β)LINEX at v = −2 takes the smallest values for n = 10, 70 and 100. The width of the BCI for ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for most values of n (see Table 10).
● At β=5, the ERs for DCRRE estimates get the smallest values at γR(β)= 0.0919312, for most values of n under most selected loss functions (see for example Figures 9 and 10).
● The ER of ˆγR,BE(β)LINEX at v = −2 gets the smallest values for all values of n. While, the width of the BCI for ˆγR,BE(β)LINEX at v = 2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for all n (see Table 11).
● The ER of ˆγR,BE(β)LINEX at v = −2 gets the smallest values for all values of n. While, the width of the BCI for ˆγR,BE(β)SE is the shortest compared to the width of the BCI in case of the LINEX and PRE loss functions for all n expect at n = 10 (see Table 12).
● Under SE loss function, the ERs for DCRRE estimates get the smallest values at γR(β)= 0.04176352, for most values of n (see for example Figures 11 and 12).
● The ER of ˆγR,BE(β)LINEX at v = 2 gets the smallest values for n = 10, 70 and 100. While, the width of the BCI for ˆγR,BE(β)LINEX at v = 2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for all n (see Table 13).
● The ER of ˆγR,BE(β)LINEX at v = 2 gets the smallest values for all values of n. While, the width of the BCI for ˆγR,BE(β)LINEX at v = −2 is the shortest compared to the width of the BCI in case of the SE and PRE loss functions for all n (see Table 14).
5.
Application to real data
In this section, we provide an application to real data set to prove the importance and flexibility of the Lomax distribution compared with some other models. Also, the real data set can be used to illustrate the proposed methods in Sections 2 and 3.
The real data set was used by Jorgensen [41] which represent the active repair times (in hours) for airborne communication transceiver. The data are recorded as follows:
Firstly, some preliminary data analysis is performed. The histogram, scaled total time on test (TTT) plots and the log likelihood for the real dataset are presented in Figure 13. We can observe that the shape of TTT plots is convex curve for dataset, which demonstrates decreasing failure rate.
Figure 13.
Estimated PDF, TTT plots and log likelihood of Lomax distribution for active repair times data for airborne communication transceiver.
Secondly, an application to a real data set is given to demonstrate the importance and flexibility of the proposed distribution compared with other one, two and three parameters fitted distributions. The suggested distributions are the alpha power Lindley (APLi) (Dey et al. [42]), Lindley (Li) and alpha power Weibull (APW) (Nassar et al. [43]) distributions. The MLE of the unknown parameters and the corresponding standard error (SE) of Lomax distribution and the other competitive models are calculated. In order to compare the different models with the Lomax model, we use measures like, Akaike information criterion (AIC), and Bayesian information criterion (BIC), consistent AIC (CAIC), Hannan-Quinn information criterion (HQIC). Also, the Kolmogorov- Smirnov (K-S) goodness of fit test and its P value (PV) for all models are calculated. Table 15 reports the MLEs and the corresponding SEs for all models. Table 16 reports numerical values of AIC, BIC, CAIC, HQIC, K-S and PVs for the different competitive distributions. From the results, we found that the Lomax distribution takes the smallest values of the considered measures and the largest value of PV among all other competitive models, and so it could be chosen as the best model.
As anticipated, we conclude from Table 17 that the DCRRE estimates are decreasing function on time (t), that is the estimated values of the DCRRE decrease as the time t increases.
6.
Conclusions
In this paper, the maximum likelihood and Bayesian methods of estimation for dynamic cumulative residual Rényi entropy of Lomax distribution are considered. The maximum likelihood estimates and approximate confidence intervals of the DCRRE are obtained. The performance of the DCRRE estimates for Lomax distribution is investigated via simulation in terms of their mean square error, average length and coverage probability. The Bayesian estimate of the DCRRE for Lomax model is obtained by considering the gamma priors under symmetric and asymmetric loss functions. The MCMC procedure is employed to compute the Bayes estimates and the Bayesian credible intervals. The performance of the DCRRE estimates for Lomax distribution is inspected through their relative absolute bias, estimated risk and the width of the credible intervals. Application to real data is provided.
Regarding the simulation results, we conclude that the mean square error of maximum likelihood and Bayesian estimates of the DCRRE decreases as the sample size increases. Also, the average length of the DCRRE estimates decreases and coverage probability increases as the sample size increases. The ER and the width of credible intervals of the DCRRE Bayes estimates decrease. For small true values of DCRRE, the width of DCRRE of BCIs under LINEX loss function is smaller than the corresponding based on SE and PRE loss functions for all selected sample size at t = 0.5 and 1.5. For large true values of DCRRE, the BCI width of the DCRRE under LINEX loss function is smaller than the corresponding based on SE and PRE loss functions for large sample size at t = 0.5 and 1.5.
Generally, the DCRRE estimates for both ML and Bayesian methods approach the true value as the sample size increases. As the time (t) increases, both ML and Bayesian estimates of the DCRRE decrease. Bayesian estimates under LINEX loss function at v = −2 are more suitable than other estimates under SE and PRC loss functions in most of the situations.
Acknowledgments
The authors thank the reviewers for their detailed and constructive comments. This project is supported by Researchers Supporting Project number (RSP-2020/156) King Saud University, Riyadh, Saudi Arabia. The first author, therefore, gratefully acknowledges the KSU for technical and financial support.
Conflict of interest
The authors declare no conflict of interest.
References
[1]
K. Patel, H. Han, A. K. Jain, Secure face unlock: Spoof detection on smartphones, IEEE Trans. Inf. Forensics Secur., 11 (2016), 2268–2283. https://doi.org/10.1109/TIFS.2016.2578288 doi: 10.1109/TIFS.2016.2578288
[2]
A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet classification with deep convolutional neural networks, Commun. ACM, 60 (2017), 84–90. https://doi.org/10.1145/3065386 doi: 10.1145/3065386
[3]
Y. Sun, Y. Chen, X. Wang, X. Tang, Deep learning face representation by joint identification-verification, in Proceedings of the 27th International Conference on Neural Information Processing Systems, 2 (2014), 1988–1996.
[4]
A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, et al., An image is worth 16×16 words: Transformers for image recognition at scale, preprint, arXiv: 2010.11929.
[5]
Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, et al., Swin transformer: Hierarchical vision transformer using shifted windows, in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), (2021), 9992–10002. https://doi.org/10.1109/ICCV48922.2021.00986
[6]
J. Yang, Z. Lei, S. Z. Li, Learn convolutional neural network for face anti-spoofing, preprint, arXiv: 1408.5601.
[7]
Z. Xu, S. Li, W. Deng, Learning temporal features using LSTM-CNN architecture for face anti-spoofing, in 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR), (2016), 141–145. https://doi.org/10.1109/ACPR.2015.7486482
[8]
Z. Yu, C. Zhao, Z. Wang, Y. Qin, Z. Su, X. Li, et al., Searching central difference convolutional networks for face anti-spoofing, in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2020), 5294–5304. https://doi.org/10.1109/CVPR42600.2020.00534
[9]
R. Cai, Z. Li, R. Wan, H. Li, Y. Hu, A. C. Kot, Learning meta pattern for face anti-spoofing, IEEE Trans. Inf. Forensics Secur., 17 (2022), 1201–1213. https://doi.org/10.1109/TIFS.2022.3158551 doi: 10.1109/TIFS.2022.3158551
[10]
R. Cai, Z. Yu, C. Kong, H. Li, C. Chen, Y. Hu, et al., S-adapter: Generalizing vision transformer for face anti-spoofing with statistical tokens, IEEE Trans. Inf. Forensics Secur., 19 (2024), 8385–8397. https://doi.org/10.1109/TIFS.2024.3420699 doi: 10.1109/TIFS.2024.3420699
[11]
R. Cai, Y. Cui, Z. Li, Z. Yu, H. Li, Y. Hu, et al., Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less, in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), (2023), 8003–8014. https://doi.org/10.1109/ICCV51070.2023.00738
[12]
A. Liu, Z. Tan, Z. Yu, C. Zhao, J. Wan, Y. Liang, et al., FM-ViT: Flexible modal vision transformers for face anti-spoofing, IEEE Trans. Inf. Forensics Secur., 18 (2023), 4775–4786. https://doi.org/10.1109/TIFS.2023.3296330 doi: 10.1109/TIFS.2023.3296330
[13]
C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. J. Goodfellow, et al., Intriguing properties of neural networks, preprint, arXiv: 1312.6199.
[14]
A. Aldahdooh, W. Hamidouche, O. Deforges, Reveal of vision transformers robustness against adversarial attacks, preprint, arXiv: 2106.03734.
[15]
A. Kurakin, I. J. Goodfellow, S. Bengio, Adversarial examples in the physical world, preprint, arXiv: 1607.02533.
[16]
K. Eykholt, I. Evtimov, E. Fernandes, B. Li, A. Rahmati, C. Xiao, et al., Robust physical-world attacks on deep learning visual classification, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2018), 1625–1634. https://doi.org/10.1109/CVPR.2018.00175
[17]
M. Sharif, S. Bhagavatula, L. Bauer, M. K. Reiter, Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition, in Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, (2016), 1528–1540. https://doi.org/10.1145/2976749.2978392
[18]
Y. Liu, X. Chen, C. Liu, D. Song, Delving into transferable adversarial examples and black-box attacks, preprint, arXiv: 1611.02770.
[19]
W. Ma, Y. Li, X. Jia, W. Xu, Transferable adversarial attack for both vision transformers and convolutional networks via momentum integrated gradients, in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), (2023), 4607–4616. https://doi.org/10.1109/ICCV51070.2023.00427
[20]
P. Russu, A. Demontis, B. Biggio, G. Fumera, F. Roli, Secure kernel machines against evasion attacks, in Proceedings of the 2016 ACM workshop on artificial intelligence and security, (2016), 59–69. https://doi.org/10.1145/2996758.2996771
[21]
F. Tramèr, N. Papernot, I. Goodfellow, D. Boneh, P. McDaniel, The space of transferable adversarial examples, preprint, arXiv: 1704.03453.
[22]
F. Yang, Z. Chen, Using randomness to improve robustness of machine-learning models against evasion attacks, preprint, arXiv: 1808.03601.
[23]
Z. Zhou, J. Feng, Deep forest: Towards an alternative to deep neural networks, in Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, (2017), 3553–3559. https://doi.org/10.24963/ijcai.2017/497
[24]
J. Maatta, A. Hadid, M. Pietikäinen, Face spoofing detection from single images using micro-texture analysis, in 2011 International Joint Conference on Biometrics (IJCB), (2011), 1–7. https://doi.org/10.1109/IJCB.2011.6117510
[25]
J. Yang, Z. Lei, S. Liao, S. Z. Li, Face liveness detection with component dependent descriptor, in 2013 International Conference on Biometrics (ICB), (2013), 1–6. https://doi.org/10.1109/ICB.2013.6612955
[26]
R. Nosaka, Y. Ohkawa, K. Fukui, Feature extraction based on co-occurrence of adjacent local binary patterns, in Advances in Image and Video Technology, 7088 (2011), 82–91. https://doi.org/10.1007/978-3-642-25346-1_8
[27]
I. Chingovska, A. Anjos, S. Marcel, On the effectiveness of local binary patterns in face anti-spoofing, in 2012 BIOSIG-Proceedings of the International Conference of Biometrics Special Interest Group (BIOSIG), (2012), 1–7.
[28]
Y. Atoum, Y. Liu, A. Jourabloo, X. Liu, Face anti-spoofing using patch and depth-based CNNs, in 2017 IEEE International Joint Conference on Biometrics (IJCB), (2017), 319–328. https://doi.org/10.1109/BTAS.2017.8272713
[29]
X. Tan, Y. Li, J. Liu, L. Jiang, Face liveness detection from a single image with sparse low rank bilinear discriminative model, in Computer Vision-ECCV 2010, 6316 (2010), 504–517. https://doi.org/10.1007/978-3-642-15567-3_37
[30]
D. Gragnaniello, G. Poggi, C. Sansone, L. Verdoliva, An investigation of local descriptors for biometric spoofing detection, IEEE Trans. Inf. Forensics Secur., 10 (2015), 849–863. https://doi.org/10.1109/TIFS.2015.2404294 doi: 10.1109/TIFS.2015.2404294
[31]
Z. Boulkenafet, J. Komulainen, A. Hadid, Face antispoofing using speeded-up robust features and fisher vector encoding, IEEE Signal Process. Lett., 24 (2017), 141–145. https://doi.org/10.1109/LSP.2016.2630740 doi: 10.1109/LSP.2016.2630740
[32]
T. D. F. Pereira, A. Anjos, J. M. D. Martino, S. Marcel, LBP-TOP based countermeasure against face spoofing attacks, in Computer Vision-ACCV 2012 Workshops, 7728 (2012), 121–132. https://doi.org/10.1007/978-3-642-37410-4_11
[33]
S. R. Arashloo, J. Kittler, W. Christmas, Face spoofing detection based on multiple descriptor fusion using multiscale dynamic binarized statistical image features, IEEE Trans. Inf. Forensics Secur., 10 (2015), 2396–2407. https://doi.org/10.1109/TIFS.2015.2458700 doi: 10.1109/TIFS.2015.2458700
[34]
Z. Boulkenafet, J. Komulainen, A. Hadid, Face spoofing detection using colour texture analysis, IEEE Trans. Inf. Forensics Secur., 11 (2016), 1818–1830. https://doi.org/10.1109/TIFS.2016.2555286 doi: 10.1109/TIFS.2016.2555286
[35]
D. Menotti, G. Chiachia, A. Pinto, W. R. Schwartz, H. Pedrini, A. X. Falcão, et al., Deep representations for iris, face, and fingerprint spoofing detection, IEEE Trans. Inf. Forensics Secur., 10 (2015), 864–879. https://doi.org/10.1109/TIFS.2015.2398817 doi: 10.1109/TIFS.2015.2398817
[36]
H. Li, P. He, S. Wang, A. Rocha, X. Jiang, A. C. Kot, Learning generalized deep feature representation for face anti-spoofing, IEEE Trans. Inf. Forensics Secur., 13 (2018), 2639–2652. https://doi.org/10.1109/TIFS.2018.2825949 doi: 10.1109/TIFS.2018.2825949
[37]
R. Cai, H. Li, S. Wang, C. Chen, A. C. Kot, DRL-FAS: A novel framework based on deep reinforcement learning for face anti-spoofing, IEEE Trans. Inf. Forensics Secur., 16 (2020), 937–951. https://doi.org/10.1109/TIFS.2020.3026553 doi: 10.1109/TIFS.2020.3026553
[38]
W. Sun, Y. Song, C. Chen, J. Huang, A. C. Kot, Face spoofing detection based on local ternary label supervision in fully convolutional networks, IEEE Trans. Inf. Forensics Secur., 15 (2020), 3181–3196. https://doi.org/10.1109/TIFS.2020.2985530 doi: 10.1109/TIFS.2020.2985530
[39]
A. George, S. Marcel, Deep pixel-wise binary supervision for face presentation attack detection, preprint, arXiv: 1907.04047.
[40]
Z. Yu, X. Li, J. Shi, Z. Xia, G. Zhao, Revisiting pixel-wise supervision for face anti-spoofing, IEEE Trans. Biom. Behav. Identity Sci., 3 (2021), 285–295. https://doi.org/10.1109/TBIOM.2021.3065526 doi: 10.1109/TBIOM.2021.3065526
[41]
R. Shao, X. Lan, J. Li, P. C. Yuen, Multi-adversarial discriminative deep domain generalization for face presentation attack detection, in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2019), 10015–10023. https://doi.org/10.1109/CVPR.2019.01026
[42]
Y. Qin, Z. Yu, L. Yan, Z. Wang, C. Zhao, Z. Lei, Meta-teacher for face anti-spoofing, IEEE Trans. Pattern Anal. Mach. Intell., 44 (2022), 6311–6326. https://doi.org/10.1109/TPAMI.2021.3091167 doi: 10.1109/TPAMI.2021.3091167
[43]
Z. Yu, J. Wan, Y. Qin, X. Li, S. Z. Li, G. Zhao, NAS-FAS: Static-dynamic central difference network search for face anti-spoofing, IEEE Trans. Pattern Anal. Mach. Intell., 43 (2021), 3005–3023. https://doi.org/10.1109/tpami.2020.3036338 doi: 10.1109/tpami.2020.3036338
[44]
Y. Liu, J. Stehouwer, X. Liu, On disentangling spoof trace for generic face anti-spoofing, in Computer Vision-ECCV 2020, 12363 (2020), 406–422. https://doi.org/10.1007/978-3-030-58523-5_24
[45]
K. Zhang, T. Yao, J. Zhang, Y. Tai, S. Ding, J. Li, et al., Face anti-spoofing via disentangled representation learning, in Computer Vision-ECCV 2020, 12364 (2020), 641–657. https://doi.org/10.1007/978-3-030-58529-7_38
[46]
H. Wu, D. Zeng, Y. Hu, H. Shi, T. Mei, Dual spoof disentanglement generation for face anti-spoofing with depth uncertainty learning, IEEE Trans. Circuits Syst. Video Technol., 32 (2022), 4626–4638. https://doi.org/10.1109/TCSVT.2021.3133620 doi: 10.1109/TCSVT.2021.3133620
[47]
W. Yan, Y. Zeng, H. Hu, Domain adversarial disentanglement network with cross-domain synthesis for generalized face anti-spoofing, IEEE Trans. Circuits Syst. Video Technol., 32 (2022), 7033–7046. https://doi.org/10.1109/TCSVT.2022.3178723 doi: 10.1109/TCSVT.2022.3178723
[48]
Y. Jia, J. Zhang, S. Shan, X. Chen, Single-side domain generalization for face anti-spoofing, in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2020), 8481–8490. https://doi.org/10.1109/CVPR42600.2020.00851
[49]
Z. Wang, Z. Wang, Z. Yu, W. Deng, J. Li, T. Gao, et al., Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing, in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2022), 4113–4123. https://doi.org/10.1109/CVPR52688.2022.00409
[50]
A. Liu, C. Zhao, Z. Yu, J. Wan, A. Su, X. Liu, et al., Contrastive context-aware learning for 3D high-fidelity mask face presentation attack detection, IEEE Trans. Inf. Forensics Secur., 17 (2022), 2497–2507. https://doi.org/10.1109/TIFS.2022.3188149 doi: 10.1109/TIFS.2022.3188149
[51]
A. George, S. Marcel, Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks, IEEE Trans. Inf. Forensics Secur., 16 (2020), 361–375. https://doi.org/10.1109/TIFS.2020.3013214 doi: 10.1109/TIFS.2020.3013214
[52]
H. Li, W. Li, H. Cao, S. Wang, F. Huang, A. C. Kot, Unsupervised Domain Adaptation for Face Anti-Spoofing, IEEE Trans. Inf. Forensics Secur., 13 (2018), 1794–1809. https://doi.org/10.1109/TIFS.2018.2801312 doi: 10.1109/TIFS.2018.2801312
[53]
Y. Liu, Y. Chen, W. Dai, M. Gou, C. Huang, H. Xiong, Source-free domain adaptation with contrastive domain alignment and self-supervised exploration for face anti-spoofing, in Computer Vision-ECCV 2022, 13672 (2022), 511–528. https://doi.org/10.1007/978-3-031-19775-8_30
[54]
Y. Qin, C. Zhao, X. Zhu, Z. Wang, Z. Yu, T. Fu, et al., Learning meta model for zero-and few-shot face anti-spoofing, in Proceedings of the AAAI Conference on Artificial Intelligence, 34 (2020), 11916–11923. https://doi.org/10.1609/aaai.v34i07.6866
[55]
H. Huang, D. Sun, Y. Liu, W. Chu, T. Xiao, J. Yuan, et al., Adaptive transformers for robust few-shot cross-domain face anti-spoofing, in Computer Vision-ECCV 2022, 13673 (2022), 37–54. https://doi.org/10.1007/978-3-031-19778-9_3
[56]
L. Li, X. Feng, Z. Xia, X. Jiang, A. Hadid, Face spoofing detection with local binary pattern network, J. Visual Commun. Image Representation, 54 (2018), 182–192. https://doi.org/10.1016/j.jvcir.2018.05.009 doi: 10.1016/j.jvcir.2018.05.009
[57]
A. Roohi, S. Angizi, Efficient targeted bit-flip attack against the local binary pattern network, in 2022 IEEE International Symposium on Hardware Oriented Security and Trust (HOST), (2022), 89–92. https://doi.org/10.1109/HOST54066.2022.9839959
[58]
N. Bousnina, L. Zheng, M. Mikram, S. Ghouzali, K. Minaoui, Unraveling robustness of deep face anti-spoofing models against pixel attacks, Multimedia Tools Appl., 80 (2021), 7229–7246. https://doi.org/10.1007/s11042-020-10041-1 doi: 10.1007/s11042-020-10041-1
[59]
D. Deb, X. Liu, A. K. Jain, Unified detection of digital and physical face attacks, in 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), (2023) 1–8. https://doi.org/10.1109/FG57933.2023.10042500
[60]
T. K. Ho, The random subspace method for constructing decision forests, IEEE Trans. Pattern Anal. Mach. Intell., 20 (1998), 832–844. https://doi.org/10.1109/34.709601 doi: 10.1109/34.709601
F. T. Liu, M. T. Kai, Y. Yu, Z. Zhou, Spectrum of variable-random trees, J. Artif. Intell. Res., 32 (2008), 355–384.
[63]
T. Chen, C. Guestrin, XGBoost: A scalable tree boosting system, preprint, arXiv: 1603.02754.
[64]
T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell., 24 (2002), 971–987. https://doi.org/10.1109/TPAMI.2002.1017623 doi: 10.1109/TPAMI.2002.1017623
[65]
Z. Zhang, J. Yan, S. Liu, Z. Lei, D. Yi, S. Z. Li, A face antispoofing database with diverse attacks, in 2012 5th IAPR International Conference on Biometrics (ICB), (2012), 26–31. https://doi.org/10.1109/ICB.2012.6199754
[66]
P. Viola, M. Jones, Rapid object detection using a casacde of simple features, in Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, (2001). https://doi.org/10.1109/CVPR.2001.990517
[67]
K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, preprint, arXiv: 1409.1556.
[68]
K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2016), 770–778, . https://doi.org/10.1109/CVPR.2016.90
[69]
I. J. Goodfellow, J. Shlens, C. Szegedy, Explaining and harnessing adversarial examples, preprint, arXiv: 1412.6572.
[70]
Z. Sun, L. Sun, Q. Li, Investigation in spatial-temporal domain for face spoof detection, in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), (2018), 1538–1542. https://doi.org/10.1109/ICASSP.2018.8461942
[71]
T. M. Oshiro, P. S. Perez, J. A. Baranauskas, How many trees in a random forest? in Machine Learning and Data Mining in Pattern Recognition, 7376 (2012), 154–168. https://doi.org/10.1007/978-3-642-31537-4_13
[72]
J. Lu, V. E. Liong, J. Zhou, Cost-sensitive local binary feature learning for facial age estimation, IEEE Trans. Image Process., 24 (2015), 5356–5368. https://doi.org/10.1109/TIP.2015.2481327 doi: 10.1109/TIP.2015.2481327
[73]
J. Lu, V. E. Liong, J. Zhou, Deep hashing for scalable image search, IEEE Trans. Image Process., 26 (2017), 2352–2367. https://doi.org/10.1109/TIP.2017.2678163 doi: 10.1109/TIP.2017.2678163
[74]
J. Lu, V. E. Liong, J. Zhou, Simultaneous local binary feature learning and encoding for homogeneous and heterogeneous face recognition, IEEE Trans. Pattern Anal. Mach. Intell., 40 (2018), 1979–1993. https://doi.org/10.1109/TPAMI.2017.2737538 doi: 10.1109/TPAMI.2017.2737538
[75]
Y. Duan, J. Lu, J. Feng, J. Zhou, Context-aware local binary feature learning for face recognition, IEEE Trans. Pattern Anal. Mach. Intell., 40 (2018), 1139–1153. https://doi.org/10.1109/TPAMI.2017.2710183 doi: 10.1109/TPAMI.2017.2710183
This article has been cited by:
1.
Zhongzhe Ouyang, Min Lu,
Systemic Financial Risk Forecasting with Decomposition–Clustering-Ensemble Learning Approach: Evidence from China,
2024,
16,
2073-8994,
480,
10.3390/sym16040480
Yi-Chiuan Wang, Yi-hao Lai, Jyh-Lin Wu,
Asymmetries in risk spillovers between currency and stock markets: Evidence from the CoVaR-copula approach,
2024,
63,
0924-865X,
1083,
10.1007/s11156-024-01285-1
4.
Zisheng Ouyang, Zhen Chen, Xuewei Zhou, Zhongzhe Ouyang,
Imported risk in global financial markets: Evidence from cross-market connectedness,
2025,
10629408,
102374,
10.1016/j.najef.2025.102374
Rizhao Cai, Liepiao Zhang, Changsheng Chen, Yongjian Hu, Alex Kot. Learning deep forest for face anti-spoofing: An alternative to the neural network against adversarial attacks[J]. Electronic Research Archive, 2024, 32(10): 5592-5614. doi: 10.3934/era.2024259
Rizhao Cai, Liepiao Zhang, Changsheng Chen, Yongjian Hu, Alex Kot. Learning deep forest for face anti-spoofing: An alternative to the neural network against adversarial attacks[J]. Electronic Research Archive, 2024, 32(10): 5592-5614. doi: 10.3934/era.2024259
Figure 1. Examples of genuine faces and spoofing faces from the MSU USSA database [1]. Columns (a), (b), and (c) are the genuine face, display face, and printed photo face and their corresponding magnifying regions, respectively
Figure 2. The illustration of how the GSM learns representations for local information [23]. First, a sliding window with a certain stride is used to scan raw pixels. Then, all the scanned patches are fed to forests, an RF (black) and a CRF (rose). Finally, all the output results from the forests will be concatenated as the representations of the raw pixel data. For full details about the GSM, please refer to [23]
Figure 3. Illustrations of constructing multi-scale representations. (a) and (b) illustrate how the MGSM and the proposed scheme construct representations on multi-scales, respectively
Figure 4. The procedure of deep forest learning with multi-scale representations. The left part contains the LBP representations on three different scales, denoted by S1LBP, S2LBP, and S3LBP, respectively. The right part illustrates the cascading strategy. The black and red boxes are output results of each forest from the previous layer. They will be concatenated with features on different scales in different layers as the input of their next layers
Figure 5. Examples of PA. (a): Display attack: The face is in a digital display screen [27]. (b) Replay attack [65]: The face is in a video. (c) Print attack: The face is in a print photo [1]. (d) Print attack: The face is in a print photo that is tailored [52]
Figure 6. The convergence curves of experiments on the MSU database. An average of the results of the five validations is taken. The x-axis refers to the number of the cascade layer that increases along with the training process. The y-axis refers to the testing accuracy of the output of each layer