2017 IEEE 3rd International Conference on Collaboration and Internet Computing Detection of Phishing Webpages using Heterogeneous Transfer Learning Karl R. Weiss Taghi M. Khoshgoftaar Florida Atlantic University kweiss6@fau.edu Florida Atlantic University khoshgof@fau.edu attempt to trick and defraud a legitimate internet user into disclosing personal information by displaying webpages that falsely represent a genuine company or cause. The types of user information that are typically targeted by phishers include user-names, passwords, answers to security questions, social security numbers, and credit card information. The information obtained by the phishing websites is typically used for malicious purposes such as identity fraud and illegal financial transactions. In 2015, the average 10,000 employee company spent $3.7 million on phishing attacks [4]. There are a number of ways to combat phishing attacks. One method of combating phishing is through legal channels. Many countries have strict laws that prohibit phishing. Anti-phishing laws have been enacted in the United States and United Kingdom since 2006 [5]. However, phishing still remains an issue in countries where strict anti-phishing laws have been enacted. Another method to combat the phishing problem is through user education. An internet user can look for items that indicate a phishing webpage, such as an incorrect domain name in the URL or not using a secure connection (e.g. https) for financial transactions. Unfortunately, the education method requires the internet user to have some level of technical ability and knowledge. Because of this, user education is only able to address a portion of the phishing problem. A third method to combat the phishing problem is through technical solutions. One technical solution that has been deployed is the use of blacklists in security software. The inherent nature of phishing websites is they do not last for long periods of time and they change their domain names frequently. An issue with using blacklists is keeping the list current and up-to-date given the dynamic nature of the phishing websites. The other technical solution is the data mining and machine learning approach, where key features of phishing websites are discovered and used to train machine learning algorithms to detect phishing webpages. This paper is focused on machine learning solutions for detecting phishing websites. A subclass of machine learning is transfer learning. Transfer learning is used in an environment where there is limited or no labeled training data for a domain of interest (referred to as the target data). In this case, labeled data from a domain, that is similar to the domain of interest (referred to as the source data), is used to train a transfer learning algorithm. Transfer learning algorithms attempt to transfer relevant information Abstract—The detection of phishing websites using traditional machine learning methods has been demonstrated in previous studies. Traditional machine learning methods assume that the input feature space is the same between the training and testing data. There are scenarios in machine learning, where the available labeled training data has a different input feature space than the testing data. In cases where the input feature space between the testing and training data are different, traditional machine learning methods cannot be used. Heterogeneous transfer learning methods are used to transform the different input feature spaces between the testing and the training data into a unique and common set of input features. For our experiment, we construct numerous scenarios for the application of phishing website detection, where the features of the testing and training data are different. Our experiment starts with a baseline dataset for the detection of phishing websites. This baseline dataset is used to create separate training and testing datasets by splitting the features, such that the features in the training and testing data are mutually exclusive. Then, a heterogeneous transfer learning technique called Canonical Correlation Analysis is used to align the input feature space between the training and testing data. The feature aligned training and testing data is used with various traditional machine learning methods and homogeneous transfer learning methods to predict phishing websites. The performance results of the different scenarios and algorithms are reported and analyzed. Keywords–Heterogeneous transfer learning, Phishing website detection. I. I NTRODUCTION The internet has become an integral part of people’s lives over the past 20 years. In 2016, there were over 286 million internet users in the United States representing approximately 88% of the population. In the United Kingdom and Japan, internet usage represents 92% and 91% of the population, respectively [1]. As of 2017, there are 3.8 trillion internet users worldwide [2]. One reason for the popularity of the internet is the ability to conveniently engage in commerce transactions. E-commerce revenue for 2017 is estimated to be over $2.2 trillion and is predicted to grow to over $4.4 trillion by 2021 [3]. Further, the internet has become a popular vehicle for banking and stock trading. With the wide use of financial activity being conducted over the internet, data security is an increasingly important factor. A threat to data security over the internet is phishing. Phishing is an attempt to steal an internet user’s personal and/or financial information by impersonating a website of a legitimate internet entity. A dishonest internet entity will 0-7695-6303-1/17/31.00 ©2017 IEEE DOI 10.1109/CIC.2017.00034 190 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. from the source data, while suppressing the unrelated portions, to build a learner to predict target data [6]. Because the source and target data are drawn from different domains in a transfer learning environment, the data distribution between the source and target data may be different [7]. When there are distribution differences between the training and testing data, the performance of a predictive learner may be degraded [8]. Transfer learning algorithms are able to build high-performing learners from training and testing data with different data distributions. There are two classes of transfer learning, which are referred to as homogeneous transfer learning and heterogeneous transfer learning. Homogeneous transfer learning is the case where the source and target data share the same input feature space, but have different data distributions. Heterogeneous transfer learning is the case where the source and target data have a different input feature space. For more information on the details of transfer learning, refer to the survey of transfer learning [6]. A machine learning practitioner may have limited labeled datasets available to train a predictive learner for a phishing application. In previous works that address the detection of phishing websites, there is no uniform agreement on which input features to use. The unlabeled testing data of interest to the machine learning practitioner may have a different input feature space from the available labeled datasets. This could be due to features being different between the training and testing data, or the training and testing data having a different number of features. If the feature input space is not the same, then traditional machine learning methods cannot be used. To this end, we apply heterogeneous transfer learning to the application of phishing website detection. This experiment is a benefit to a machine learning practitioner, who is in search of a solution for the application of phishing website detection, where the available labeled training does not have the same input feature space as the testing data. Our experiment constructs dataset pairs with mutually exclusive features to create heterogeneous test scenarios. A heterogeneous test scenario is where the source and the target data have a different input feature space. In our experiment, the input feature space of the source and target data is first aligned (transformed to be the same) using a heterogeneous transfer learning technique called Canonical Correlation Analysis (CCA) [9]. Then, the feature aligned source and target data is used to train various homogeneous transfer learning algorithms, along with traditional machine learning algorithms. Our goal is to determine the effectiveness of using heterogeneous transfer learning for the application of phishing website prediction. We further examine the effectiveness using homogeneous transfer learning and traditional machine learning with the the transformed input feature space. To the best of our knowledge, heterogeneous transfer learning methods coupled with both traditional machine learning and homogeneous transfer learning methods have not been applied to the application of phishing website detection. The remainder of this paper is organized as follows. Section II discusses works relevant to our experiment. Section III presents the experiment design and definition. Section IV provides and discusses the test results of the experiment. Section V presents the experiment conclusions. II. R ELATED W ORK The majority of previous heterogeneous transfer learning works focus solely on aligning the input feature space of the source and target data. These works make an assumption that there is no further distribution difference between the source and target data after the feature alignment step. However, there are limited works that address both input feature alignment and reducing distribution differences between the source and target data. The work by Zhou et al. [10] first learns an asymmetric transformation from the target data to the source data to align the input feature space. Next, the algorithm discovers a common latent feature space using the transformed data to reduce the distribution differences between the target data and the source data. This portion of the algorithm is implemented using a deep learning method. Lastly, the common latent feature space is used to train a learner. The algorithm by Zhou et al. [10] is used in the application of multiple language text sentiment classification. The heterogeneous transfer learning work by Yi-Ren et al. [9] employs a domain adaptation technique to reduce distribution differences between the source and target data, and also uses the CCA algorithm to resolve the input feature space differences. A kernelized version of the CCA algorithm is also used to derive a nonlinear correlation subspace which aligns the input feature space between the source and target data. The aligned feature space is used to train a Support Vector Machine (SVM) model that is modified with a domain adaptation technique. This method is tested using the applications of image recognition and text-image classification. There have been a number of works that have addressed the topic of phishing detection through traditional machine learning techniques. The paper by Aburrous et al. [11] defines 27 unique features to detect e-banking phishing. These 27 features are assigned one of three values of Legitimate, Genuine, or Doubtful. The experiment focuses on the comparison of six different associative classification algorithms. In the work by Pan and Ding [12], six unique features are defined to predict the existence of a phishing webpage. A Support Vector Machine is used as the traditional machine learning model. The accuracy level obtained is below other comparable studies, which could be due to the small number of features used. The work by Fette et al. [13] uses Random Forest, Support Vector Machine, Decision Tree, and Naive Bayes models to predict phishing emails. Ten features are constructed from the emails collected for the experiment. The work by Mohammad et al. [14] uses 17 unique features to detect phishing websites. This method uses a self-structuring neural network to perform the classification process. Our experiment applies a heterogeneous transfer learning technique, coupled with homogeneous transfer learning and 191 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. the URL has one subdomains it is labeled as Legitimate, if it is has two subdomains it is labeled as Suspicious, otherwise it is labeled as Phishing. • Using Secure Socket Layer (Feature H): Using a Secure Socket Layer (e.g. https) gives the impression of legitimacy to a webpage. This false sense of security is used by phishers to trick users. There are three criteria used for this feature: use of https, use of a trusted certificate authority, and length of the issued certificate. A list of trusted certificate authorities is obtained for this feature evaluation. If using http (non-secure), then the feature is labeled as Legitimate. If using https and the certificate authority is trusted and the age of the certificate is less than one year, then the feature is labeled as Legitimate. If using https and the certificate authority is not trusted, then the feature is labeled as Suspicious, otherwise it is labeled as Phishing. • Domain name registration length (Feature I): Phishing websites typically have a short life span, therefore their domain names are only registered for the minimum amount of time. If the domain name registration expires in less than one year, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using a favicon (Feature J): A favicon is a small icon image that is shown with the address or title of the webpage. If the favicon is loaded from a domain different from what is shown in the address bar, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using a non-standard port (Feature K): A list of typical standard port usage is obtained for this feature. If a request is made to use a non-standard port, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using an https token in the domain name (Feature L): Phishers may try to trick users by adding https to the domain name of a URL. If the domain name includes https, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using request URL (Feature M): Legitimate websites typically request image and video files from the same domain name as the webpage. If less than 22% of requests are made to different domains, then it is labeled as Legitimate. If between 22% and 61% of the requests are made to different domains, then it is labeled as Suspicious. If more than 61% of the requests are made to different domains, then it is labeled as Phishing. • Anchor requests (Feature N): An anchor is an html tag defined as <a>. Here, the percentage that an anchor tag makes a request from a domain name different from the webpage domain name is tabulated combined with the percentage that an anchor tag does not point to any domain name. If the percentage is less than 31%, then it is labeled as Legitimate. If the percentage is between 31% and 67%, then it is labeled as Suspicious. If the percentage is above 67%, then it is labeled as Phishing. • Use of link tags (Feature O): The <meta>, <link>, and <script> tags in the html code are examined to find out how many make requests to a domain name that is different from the webpage. If the percentage is less than 17%, then it is labeled as Legitimate. If the percentage is between 17% and traditional machine learning models to the application of detecting phishing websites. Our experiment parameters are unique from the previous related work, and to the best of our knowledge have not been previously investigated. III. E XPERIMENT D ESIGN AND D EFINITION The following section defines the datasets, learners, and procedures used to perform our experiment. Parameter values for the experiment are provided in addition to defining the processes needed to ensure the experiment results are reproducible and reliable. A. Base Dataset The dataset used for our experiment is called “Phishing Websites” which is sourced from the UCI Machine Learning Repository [15]. There are 11,055 instances in this dataset with approximately 55% of the instances labeled as “Phishing” and the other 45% labeled as “Legitimate.” There are 30 independent features in the Phishing Websites dataset. The features are assigned two values of Legitimate or Phishing, or three values of Legitimate, Suspicious or Phishing. In this section, we define the 30 features used in this paper. For further details on these features, refer to “Phishing Websites” at the UCI Machine Learning Repository [15]. • Using an IP address (Feature A): If the URL contains an actual numbered IP address (e.g. http://11.96.10.20/a.html) as opposed to a domain name, then the feature is labeled as Phishing, otherwise it is labeled as Legitimate. • Using a long URL (Feature B): Phishers sometimes try to hide a suspicious part of the URL by creating long URLs. If the URL is less than 54 characters, the feature is labeled as Legitimate, if the URL is between 54 and 75 characters the feature is labeled as Suspicious, otherwise the feature is labeled as Phishing. • Using URL Shortening Services (Feature C): A shortening service uses a URL name (usually a small URL) that is different from the actual URL to redirect the webpage. If a shortening service is used the feature is labeled as Phishing, otherwise it is Legitimate. • Using “@” symbol in URL (Feature D): The characters preceding an “@” symbol in a URL are ignored by the browser. This method is used in an attempt to trick users into thinking the URL is different from what it actually is. Using an “@” symbol in the URL is considered Phishing, otherwise it is Legitimate. • Using “//” in URL to redirect (Feature E): If a “//” is used in a URL to redirect to another website, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using “-” in the domain name (Feature F): A dash symbol (“-”) can be used in a domain name to fool a user into thinking the domain name is from a legitimate company. If a “-” is used in the domain name, then it is labeled as Phishing, otherwise it is Legitimate. • Using multiple subdomains in a URL (Feature G): A typical URL has one subdomain. This subdomain can be “www” (e.g. as in www.ibm.com) or some other domain defined name. If 192 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. • Google index (Feature Z3): Most phishing webpages are not indexed by Google. If a webpage is indexed by Google, then it is labeled as Legitimate, otherwise it is labeled as Phishing. • Number of links (Feature Z4): Phishing webpages typically do not have links from other webpages pointing to their webpage. If the number of links pointing to the webpage is zero, then it is labeled as Phishing. If the number of links pointing to the webpage is between one and two, then it is labeled as Suspicious. If the number of links pointing to the webpage is greater than two, then it is labeled as Legitimate. • Phishing reports (Feature Z5): The statistics from PhishTank and StopBadware are used for this feature. If the host for the webpage has an IP address or domain name listed in PhishTank or StopBadware list, then it is labeled Phishing, otherwise it is labeled Legitimate. 81%, then it is labeled as Suspicious. If the percentage is above 81%, then it is labeled as Phishing. • Use of server form handler (Feature P): In legitimate websites, the server form handler typically links to a domain name that is the same as the webpage. If the server form handler is blank or empty, then it is labeled as Phishing. If the server form handler links to a domain name that is different from the webpage, then it is labeled as Suspicious. For all other cases, it is labeled as Legitimate. • Submitting user information to email (Feature Q): Legitimate websites typically will not use the mailto or mail() functions to send user information. If a website uses the mailto or mail() functions, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Abnormal URL (Feature R): If the host-name (obtained from the WHOIS database) is contained in the URL, then it is labeled as Legitimate, otherwise it is labeled as Phishing. • Number of redirects (Feature S): Legitimate websites are redirected, on average, less than Phishing websites. If the number of redirections is zero or one, then it is labeled as Legitimate. If the number of redirections is between two and four, then it is labeled as Suspicious. If the number of redirections is greater than four, than it is labeled as Phishing. • Status bar change (Feature T): Phishing websites can use Javascript to display an erroneous URL in the status bar. If the mouseover event in a webpage changes the status bar, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Disable mouse right-click (Feature U): Phishing websites can disable the mouse right-click function so users cannot view the webpage source code. If the mouse right-click is disabled, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Using pop-up windows (Feature V): Phishing webpages can ask users to submit personal information on pop-up windows, whereas this activity is rare for Legitimate webpages. If there are text fields in a pop-up window, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Use of iframe tag (Feature W): If a webpage uses at least one iframe tag, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Age of domain (Feature X): Phishing websites do not last for long. The age of the domain is obtained from the WHOIS database. If the age of the domain is less than six months, then it is labeled as Phishing, otherwise it is labeled as Legitimate. • DNS record (Feature Y): If the DNS record is not found or is empty (per the WHOIS database) then it is labeled as Phishing, otherwise it is labeled as Legitimate. • Website traffic (Feature Z1): Phishing websites typically do not generate large amounts of traffic. If the website traffic rank is in the top 100,000, then it is labeled as Legitimate. If the website traffic is not in the top 100,000 (but still ranked), then it is labeled as Suspicious. If the website is not ranked, then it is labeled as Phishing. • Page rank (Feature Z2): Phishing webpages typically have a very low page rank. If the page rank of a webpage is less than 0.2, then it is labeled as Phishing, otherwise it is labeled as Legitimate. B. Dataset Construction From the Base Dataset of Phishing Websites described above, we divide both the instances and the features to create source and target dataset pairs for our experiment. To aid in the feature division process, the relevance of the 30 features is determined by using the Chi-Square measure [16]. The following is the feature ranking from most relevant to least relevant: H, N, Z1, G, F, O, M, I, P, Z3, X, Z2, A, Z5, Z4, Y, B, C, R, D, T, L, E, K, S, Q, U, W, J, V. There are three datasets pairs (six datasets in total) that are created from the Phishing Websites Base Dataset. A dataset pair represents a source and target data pair. Each dataset within the dataset pair contains 5,500 instances which are randomly selected from the Base Dataset. Each dataset pair is constructed to contain mutually exclusive features from the Base Dataset. The first dataset pair is created by evenly distributing the features by order of relevance. For this case, the individual datasets are referred to as 15AF and 15BF, where each dataset contains 15 unique features. The next dataset pair, referred to as 10AF and 10BF, is created by eliminating the five least relevant features from the 15AF and 15BF datasets. The 10AF and 10BF datasets each have 10 unique features. The last dataset pair, referred to as 9F and 21F, is created by assigning the nine features used from the work of Abdelhamid et al. [17] to the 9F dataset. Then, the remaining 21 features are assigned to the 21F dataset. The 9F dataset has a higher concentration of high-ranked relevant features as compared to the 21F dataset. The class balance for each dataset is maintained at approximately 55% for the “Phishing” class. The dataset pairs are summarized in Table I. C. Algorithms Tested The algorithm structure used for our experiment is depicted in Figure 1. For the “Heterogeneous Transfer Learning” section, we use the Canonical Correlation Analysis algorithm as developed by Yi-Ren et al. [9]. Source and target data with different feature spaces are input to the CCA algorithm. The output (“Feature Aligned Data”) is a transformed source and target data with the same feature spaces. This information 193 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. TABLE I D EFINITION OF DATASET PAIRS Dataset Name 15AF 15BF 10AF 10BF 9F 21F the Geodesic Flow Kernel (GFK) algorithm [22], the Joint Domain Adaptation (JDA) algorithm [23], the Transfer Joint Matching (TJM) algorithm [24], the Graph Co-Regularization Transfer Learning (GTL) algorithm [25], the Transfer Kernel Learning (TKL) algorithm [26], and the Transfer Component Analysis (TCA) algorithm [27]. These seven homogeneous transfer learning algorithms require labeled source data and unlabeled target data. All seven of the transfer learning algorithms are open source implementations written in the MATLAB programming language. The four traditional machine learning algorithms used in our experiment are SVM [28], 5Nearest Neighbors (5NN) [29], Logistic Regression (LR) [30], and Random Forest (RF) [31]. The ARTL algorithm simultaneously resolves marginal and conditional distribution differences, while learning the final classification model. Pseudo labels are discovered for the unlabeled target data in an attempt to resolve the conditional distribution differences. The pseudo labels are found using an L2-regularized Logistic Regression model trained with the labeled source data. The distribution alignment and learning process are integrated into a Regularized Least Squares classification model. The GFK algorithm uses a feature reduction process to reduce the marginal distribution differences between the source and target data. The domain adaptation process further employs a geodesic flow kernel built with the reduced input feature space. The geodesic flow kernel projects a large number of subspaces onto the geodesic flow curve. The geodesic flow curve creates a smoothing function for the geometric and statistical properties between the target and source data. The GFK algorithm then finds domain-invariant subspaces on the geodesic flow curve to further reduce the marginal distribution differences between the target and source data. These domaininvariant subspaces are used to train a 5NN traditional learner. The JDA algorithm simultaneously corrects for the marginal and conditional distribution differences between the source and target data. A modified Principal Component Analysis (PCA) step is used for dimensionality reduction. This modified PCA algorithm is integrated with the Maximum Mean Discrepancy distance measure, which is used to resolve the marginal distribution differences. The labeled source data is used to train a classifier that generates pseudo labels, which are used in the conditional distribution correction process. An SVM traditional learner is trained with the distribution aligned data. The TJM algorithm performs a domain adaptation process by first reducing the marginal distribution between the source and target data. Using PCA, the TJM algorithm performs a dimensionality reduction step. Next, a feature matching process is employed using the Maximum Mean Discrepancy measure in a Reproducing Kernel Hilbert Space. An instance reweighing procedure is used to further reduce the distribution differences between the source and target. Irrelevant instances are eliminated during this step. The modified source data is used to train a 5NN traditional learner. The GTL algorithm uses a Nonnegative Matrix Factorization (NMF) [32] technique to reduce the marginal distribution Features C,E,L,M,N,O,Q,R,U,V,X,Z1,Z3,Z4,Z5 A,B,D,F,G,H,I,J,K,P,S,T,W,Y,Z2 C,M,N,O,R,X,Z1,Z3,Z4,Z5 A,B,D,F,G,H,I,P,Y,Z2 A,B,H,M,N,P,V,X,Z1 C,D,E,F,G,I,J,K,L,O,Q,R,S,T,U,W,Y,Z2,Z3,Z4,Z5 (along with the class labels for the source data) is input to the “Learning Algorithm” section. The Learning Algorithm section uses either a traditional machine learning model or a homogeneous transfer learning model. Source Features Target Features Heterogeneous Transfer Learning Feature Aligned Data Learning Algorithm Classified Output Fig. 1. Block diagram of the experiment algorithm structure 1) Heterogeneous Transfer Learning: Canonical Correlation Analysis is a process of discovering basis vectors for two different sets of variables. The CCA process attempts to mutually maximize the correlation between the projection of the variables onto the basis vectors. The linear transformation that is found by the CCA process creates a pair of transformed variable sets that have maximized correlation [18]. After the CCA transformation, the source and target input feature spaces are the same. For complete technical details of the algorithm refer to Yi-Ren et al. [9]. The CCA algorithm does not require class labels to perform the feature transformation. At the input to the CCA algorithm, we create an additional input feature for the source data, that directly represents the source class label. The target feature space remains unchanged. The addition of the source class label to the source input feature space adds further information during the transformation process and is found to have overall better performance in our testing as compared to not including this extra feature. The CCA algorithm used in our experiment is an open source implementation written in the MATLAB [19] programming language. 2) Learning Algorithms: There are seven state-of-the-art homogeneous transfer learning algorithms [20] and four traditional learning algorithms used as the Learning Algorithm block in our experiment. The seven homogeneous transfer learning methods used in our experiment include the Adaptation Regularization Transfer Learning (ARTL) algorithm [21], 194 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. similarity function to determine the nearest neighbors. The 5NN algorithm is implemented in MATLAB. differences between the target and source data. The NMF step finds common latent features through collective matrix factorization, while preserving the statistical and geometric properties of the target and source data. A graph coregularization process is integrated into the NMF technique to further refine the common features. The NMF process is used as the underlying learning model for the GTL algorithm. The TKL algorithm uses a Reproducing Kernel Hilbert Space to align the source and target marginal data distributions differences. At the same time, it learns a domain-invariant kernel. These extrapolated kernel spaces use the Nyström approximation error to represent the distribution differences. Numerous spectral kernels are created by using a target eigensystem technique on the source data. The final kernel is selected from the spectral kernel that most closely matches the target kernel. The data from the domain-invariant kernel is used to train an SVM traditional learner. The TCA algorithm attempts to find common latent features that have similar marginal distributions across the target and source data. The latent features are discovered using a Reproducing Kernel Hilbert Space. The Maximum Mean Discrepancy measure is used to determine the marginal distribution differences. After the latent features with reduced marginal distribution differences are found, an SVM model is trained as the traditional learner. The LR algorithm is a statistical learning method, that is used for binary classification. The LR model is found using an algebraic equation of summed terms, where each input feature is represented by a single term with a corresponding coefficient. The result of the algebraic equation is the probability of one of the output classes. The MATLAB command “glmfit” is used to implement the LR model. Default parameters are used with the exception of the distribution parameter being set to “binomial” and the link parameter being set to “logit.” The SVM model is an algorithm that finds the maximum margin linear or non-linear hyperplane within an instance space, which provides the greatest separation between the classes. The nearest instances to the hyperplane form the support vectors. In our experiment, the LIBSVM software [33] is used with a Gaussian radial basis function kernel. The SVM model is run in a MATLAB environment. Default parameters are used with the exception of the “C” parameter, which is uniquely set per dataset. The RF model is an ensemble-based prediction method, where the ensemble is comprised of N unpruned decision trees. The training process for each of the N decision trees creates random datasets using sampling with replacement. This training process ensures low model bias and decreased performance variance. The classification result is found by the majority vote (or average) of the ensemble of decision trees. The MATLAB command “TreeBagger” is used to implement the RF model. The NumTrees parameter (N) is set to 100. Settings greater than 100 did not produce better results. The 5NN learning model predicts a test instance by a majority vote of the five nearest training instances to that particular test instance. Euclidean distance is used as the D. Experiment Test Definition and Parameter Setting For the tests performed in our experiment, the source and target data each contain 5,500 instances. For our experiment, we assume a machine learning application environment with similar negative and positive class importance. The positive class (the Phishing class) balance for both the source and target data is approximately 55%. For all tests, the classifier’s discriminative threshold setting is set to a default of equal cost for a false negative and a false positive. Classification accuracy is selected as the performance measure [34]. The homogeneous transfer learning algorithms in our experiment have parameter settings that impact performance. The algorithm parameters are individually set as a function of the dataset used. The parameters used for our experiment are discovered by testing different parameter combinations for each dataset. The original algorithm manuscripts are used as a reference to determine the best range of values to search from. The source and target data are drawn from the same dataset (domain) during the parameter setting process. The algorithm parameters are summarized in Table IV (located in the Appendix). IV. D ISCUSSION OF R ESULTS This section examines the results of the CCA algorithm coupled with seven different homogeneous transfer learning algorithms and four traditional learning algorithms, when tested with the defined dataset pairs. Table II shows the accuracy results of the traditional learners using 5-fold cross-validation for each of the six datasets defined in Section III-B. Table II serves as an upper bound reference on the performance results for this experiment. The average of all the results from Table II is 87.5%. TABLE II 5- FOLD CROSS - VALIDATION FOR EACH INDIVIDUAL DATASET Algorithm SVM 5NN LR RF 15AF 89.7% 87.0% 88.1% 90.0% 15BF 90.0% 86.8% 87.5% 90.0% 10AF 88.6% 86.6% 87.6% 88.6% 10BF 90.4% 87.7% 88.8% 90.5% 9F 92.2% 90.8% 91.4% 92.5% 21F 82.0% 78.2% 74.5% 81.0% Three dataset pairs are initially defined in Section III-B. The source and target are reversed for each dataset pair to create a total of six dataset pairs. Table III shows the results for the six dataset pair combinations that are tested. The first seven algorithms listed are homogeneous transfer learning algorithms and the last four algorithms are traditional machine learning algorithms. A dataset pair is identified as the source dataset followed by the target dataset (e.g. source/target). From Table III, it is observed that both the SVM and TCA algorithms show no predictive capabilities for all of the six dataset pairs. The TCA transfer learning algorithm uses an SVM model as its base traditional learner. However, the TKL and JDA transfer learning algorithms also use an 195 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. TABLE III C LASSIFICATION ACCURACY FOR DIFFERENT DATASET PAIRS AND ALGORITHMS USING THE CCA ALGORITHM Algorithms ARTL GFK JDA TJM GTL TKL TCA SVM 5NN LR RF Average 15AF/15BF 88.9% 79.9% 88.1% 88.4% 85.9% 88.6% 45.3% 44.0% 75.9% 88.7% 88.7% 78.4% 15BF/15AF 86.1% 73.4% 79.8% 86.3% 86.3% 85.3% 55.4% 55.4% 77.8% 87.1% 87.1% 78.2% Source Dataset / Target Dataset 10AF/10BF 10BF/10AF 9F/21F 89.0% 85.7% 71.6% 82.6% 78.9% 63.7% 88.5% 81.8% 73.6% 88.2% 84.5% 74.3% 87.8% 86.2% 71.6% 88.9% 85.7% 71.6% 44.0% 55.4% 56.0% 44.0% 55.4% 44.0% 84.4% 79.3% 63.3% 88.7% 87.0% 72.2% 88.9% 86.9% 72.0% 79.5% 78.8% 66.7% SVM model as its base traditional learner. Both the TKL and JDA algorithms demonstrate good predictive capabilities over the six dataset pairs tested. Therefore, the poor performance of the TCA algorithm cannot be directly linked to the poor performance of the SVM algorithm. With the exception of the SVM and TCA algorithms, all other algorithms demonstrate good predictive capabilities over the six dataset pairs. With respect to the highest performing algorithms for each of the dataset pairs, the ARTL algorithm has the highest accuracy in three cases, the TJM algorithm in one case, the LR algorithm in three cases, and the RF algorithm in one case. The highest performing algorithm is displayed in bold text for each column in Table III. The last column in Table III shows the average accuracy for each algorithm over the six dataset pairs. The two top performing models are the RF and LR algorithms. Most notably, nine of the eleven algorithms are at or above 77% accuracy, which is within 10% of the average cross-validation results from Table II. Further, six of the algorithms are within 3% of the of the average cross-validation results. From Table III, it is observed that the homogeneous transfer learning algorithms do not demonstrate a performance advantage over the traditional learning algorithms. Both the transfer learning and traditional learning algorithms have an example of poor performance (e.g. TCA and SVM, respectively). Also, both the transfer learning and traditional learning algorithms have examples of high performers (e.g. ARTL and TJM for transfer learning, and LR and RF for traditional machine learning). The 10AF and 10BF datasets are derived by removing the five least relevant features from both the 15AF and 15BF datasets, respectively. We want to understand the impact of the five extra features on the CCA transformation. From Table III, both the 10AF/10BF and 10BF/10AF average results are marginally higher than the 15AF/15BF and15BF/15AF average results. The removal of the five least relevant features has a minimal impact on the results. The 9F dataset is constructed with a higher concentration of relevant features as compared to the 21F dataset. This is evidenced by the individual dataset cross-validation results from Table II, where the 9F dataset has at least a 10% accuracy 21F/9F 91.0% 83.5% 88.4% 90.5% 89.6% 90.6% 47.3% 44.6% 83.5% 91.0% 90.9% 81.0% Average 85.3% 77.0% 83.4% 85.4% 84.6% 85.1% 50.6% 47.9% 77.4% 85.8% 85.8% 77.1% advantage over the 21F dataset for each learner. When using the 9F and 21F dataset pair with the CCA algorithm, Table III shows that the 21F/9F dataset pair has an average accuracy advantage of approximately 14% over the 9F/21F dataset pair. Using the dataset with the least relevant features as the target (e.g. 21F) has a negative impact of the results as compared to using that dataset as the source. V. C ONCLUSION Our experiment performs heterogeneous transfer learning tests on the application of predicting phishing webpages. We construct six dataset pairs from a Base Dataset, where each dataset pair contains mutually exclusive features. A heterogeneous transfer learning technique called Canonical Correlation Analysis is used to align the features in the dataset pair. Then, seven homogeneous transfer learning and four traditional machine learning algorithms are used to learn and predict the transformed input feature space. We have demonstrated that nine of the eleven algorithms show good predictive capabilities for detecting phishing webpages, with two algorithms showing no predictive capabilities. The homogeneous transfer learning algorithms did not have a performance advantage over the traditional machine learning algorithms. This investigation is a benefit to a machine learning practitioner, who is in search of a solution for the application of phishing website detection, where the available labeled training does not have the same input feature space as the testing data. We are not aware of other works that have applied heterogeneous transfer learning methods coupled with both homogeneous transfer learning and traditional machine learning methods to the application of phishing website detection. Areas of future work include expanding the constructed dataset pairs to include a wider range of feature differences and testing different traditional machine learners. ACKNOWLEDGMENT The authors gratefully acknowledge partial support by the National Science Foundation, under grant number CNS1427536. Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the author(s) 196 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply. and do not necessarily reflect the views of the National Science Foundation. [22] B. Gong, Y. Shi, F. Sha, and K. Grauman, “Geodesic flow kernel for unsupervised domain adaptation,” Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2066– 2073, 2012. [23] M. Long, J. Wang, G. Ding, J. Sun, and P. S. Yu, “Transfer feature learning with joint distribution adaptation,” in 2013 IEEE International Conference on Computer Vision, 2013, pp. 2200–2207. [24] M. Long, J. Wang, G. Ding, J. Sun, and P. Yu, “Transfer joint matching for unsupervised domain adaptation,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1410–1417. [25] M. Long, J. Wang, G. Ding, D. Shen, and Q. Yang, “Transfer learning with graph co-regularization,” IEEE Transactions on Knowledge and Data Engineering, vol. 26, no. 7, pp. 1805–1818, 2014. [26] M. Long, J. Wang, J. Sun, and P. S. Yu, “Domain invariant transfer kernel learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 27, no. 6, pp. 1519–1532, 2015. [27] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation via transfer component analysis,” IEEE Transactions on Neural Networks, vol. 22, no. 2, pp. 199–210, 2011. [28] B. Schölkopf, C. J. C. Burges, and A. J. Smola, Eds., Advances in Kernel Methods: Support Vector Learning. Cambridge, MA, USA: MIT Press, 1999. [29] D. W. Aha, Ed., Lazy Learning. Norwell, MA, USA: Kluwer Academic Publishers, 1997. [30] T. M. Khoshgoftaar and E. B. Allen, “Logistic regression modeling of software quality,” International Journal of Reliability, Quality and Safety Engineering, vol. 6, no. 04, pp. 303–317, 1999. [31] T. M. Khoshgoftaar, M. Golawala, and J. Van Hulse, “An empirical study of learning from imbalanced data using random forest,” in 19th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2007), vol. 2. IEEE, 2007, pp. 310–317. [32] D. D. Lee and H. S. Seung, “Algorithms for non-negative matrix factorization,” in In NIPS. MIT Press, 2000, pp. 556–562. [33] C. Chang and C. Lin, “LIBSVM – A Library for Support Vector Machines,” http://www.csie.ntu.edu.tw/∼cjlin/libsvm, [Online; accessed 2-Sept-2017]. [34] K. R. Weiss and T. M. Khoshgoftaar, “Analysis of transfer learning performance measures,” in Information Reuse and Integration (IRI), 2017 IEEE International Conference on, 2017, pp. 338–344. R EFERENCES [1] “Internet users by country 2016,” [Online; accessed 2-Sept-2017]. [Online]. Available: http://www.internetlivestats.com/internet-users-bycountry [2] “Internet growth statistics 1995 to 2017,” [Online; accessed 2-Sept-2017]. [Online]. Available: http://www.internetworldstats.com/emarketing.htm [3] “Global retail e-commerce market size,” [Online; accessed 2-Sept-2017]. [Online]. Available: https://www.statista.com/statistics/379046/worldwide-retail-ecommerce-sales [4] Maria Korolov, “Phishing cost for large businesses,” [Online; accessed 2-Sept-2017]. [Online]. Available: https://www.csoonline.com/article/2975807/cyber-attacksespionage/phishing-is-a-37-million-annual-cost-for-average-largecompany.html [5] R. M. Mohammad, F. A. Thabtah, and L. McCluskey, “Predicting phishing websites based on self-structuring neural network.” Neural Computing and Applications, vol. 25, pp. 443–458, 2014. [6] K. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of transfer learning,” Journal of Big Data, vol. 3, no. 1, pp. 1–40, 2016. [7] K. R. Weiss, T. M. Khoshgoftaar, and O. Rehman, “Designing a testing framework for transfer learning algorithms,” in Information Reuse and Integration (IRI), 2016 IEEE International Conference on. IEEE, 2016, pp. 152–159. [8] H. Shimodaira, “Improving predictive inference under covariate shift by weighting the log-likelihood function,” Journal of Statistical Planning and Inference, vol. 90, no. 2, pp. 227–244, 2000. [9] Y. Yeh, C. Huang, and Y. F. Wang, “Heterogeneous domain adaptation and classification by exploiting the correlation subspace,” IEEE Trans. Image Processing, vol. 23, no. 5, pp. 2009–2018, 2014. [10] J. T. Zhou, S. J. Pan, I. W. Tsang, and Y. Yan, “Hybrid heterogeneous transfer learning through deep learning,” in Proceedings of the TwentyEighth AAAI Conference on Artificial Intelligence, ser. AAAI’14. AAAI Press, 2014, pp. 2213–2219. [11] M. Aburrous, M. A. Hossain, K. Dahal, and F. Thabtah, “Predicting phishing websites using classification mining techniques with experimental case studies,” in 2010 Seventh International Conference on Information Technology: New Generations, April 2010, pp. 176–181. [12] Y. Pan and X. Ding, “Anomaly based web phishing page detection,” in 2006 22nd Annual Computer Security Applications Conference (ACSAC’06), Dec 2006, pp. 381–392. [13] I. Fette, N. Sadeh, and A. Tomasic, “Learning to detect phishing emails,” in Proceedings of the 16th International Conference on World Wide Web, ser. WWW ’07. ACM, 2007, pp. 649–656. [14] R. M. Mohammad, F. A. Thabtah, and L. McCluskey, “Predicting phishing websites based on self-structuring neural network,” Neural Computing and Applications, vol. 25, no. 2, pp. 443–458, 2014. [15] M. Lichman, “UCI machine learning repository,” 2013, [Online; accessed 2-Sept-2017]. [Online]. Available: http://archive.ics.uci.edu/ml [16] I. H. Witten and E. Frank, Data Mining:Practical Machine Learning Tools and Techniques, 3rd ed. Morgan Kaufmann, 2011. [17] N. Abdelhamid, A. Ayesh, and F. Thabtah, “Phishing detection based associative classification data mining,” vol. 41, p. 59485959, 10 2014. [18] D. R. Hardoon, S. R. Szedmak, and J. R. Shawe-taylor, “Canonical correlation analysis: An overview with application to learning methods,” Neural Comput., vol. 16, no. 12, pp. 2639–2664, Dec. 2004. [19] The MathWorks Inc, “Matlab and statistics toolbox release 2016a,” [Online; accessed 2-Sept-2017]. [Online]. Available: https://uk.mathworks.com/products/matlab.html [20] K. R. Weiss and T. M. Khoshgoftaar, “Analysis of a transfer learning application using the transfer learning test framework,” in 23rd ISSAT International Conference on Reliability and Quality in Design, Aug 2017, pp. 151–156. [21] M. Long, J. Wang, G. Ding, S. J. Pan, and P. S. Yu, “Adaptation regularization: A general framework for transfer learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 26, no. 5, pp. 1076–1089, 2014. A PPENDIX The settable algorithm parameters for each dataset are captured in Table IV. TABLE IV S ETTABLE ALGORITHM PARAMETERS USED IN OUR EXPERIMENT Algorithm (Parameters) ARTL (p,σ,λ,γ) GFK (d) JDA (k,λ,γ,T,C) TJM (k,λ,T) GTL (p,λ,γ,σ,T) TKL (ζ,C) TCA (d,σ,C) Algorithm (Parameters) ARTL (p,σ,λ,γ) GFK (d) JDA (k,λ,γ,T,C) TJM (k,λ,T) GTL (p,λ,γ,σ,T) TKL (ζ,C) TCA (d,σ,C) 15AF 15BF 10AF 4,.2,1,.1 5 15,5,.25,5,10 12,5,10 14,1,1,100,5 .8,2 15,8,2 4,.1,5,.25 5 10,5,.1,10,1 10,10,5 5,.1,10,100,5 1.1,10 10,4,1 4,.1,5,.25 4 9,5,.1,10,1 9,10,5 5,.1,10,100,5 1.1,10 9,4,5 10BF 9F 21F 4,.1,5,.25 5 9,5,.1,10,1 9,10,5 8,.1,10,100,5 1.1,10 9,4,5 4,.2,1,.1 5 10,5,.25,5,10 10,5,10 9,1,1,100,5 .8,2 10,8,2 4,.1,5,.25 4 10,5,.1,10,1 10,10,5 5,.1,10,100,5 1.1,10 10,4,5 197 Authorized licensed use limited to: Zhengzhou University. Downloaded on May 29,2025 at 13:55:49 UTC from IEEE Xplore. Restrictions apply.
0
You can add this document to your study collection(s)
Sign in Available only to authorized usersYou can add this document to your saved list
Sign in Available only to authorized users(For complaints, use another form )