Towards robust convolutional neural networks in challenging environments
- Authors: Hossain, Md Tahmid
- Date: 2021
- Type: Text , Thesis , PhD
- Full Text:
- Description: Image classification is one of the fundamental tasks in the field of computer vision. Although Artificial Neural Network (ANN) showed a lot of promise in this field, the lack of efficient computer hardware subdued its potential to a great extent. In the early 2000s, advances in hardware coupled with better network design saw the dramatic rise of Convolutional Neural Network (CNN). Deep CNNs pushed the State-of-The-Art (SOTA) in a number of vision tasks, including image classification, object detection, and segmentation. Presently, CNNs dominate these tasks. Although CNNs exhibit impressive classification performance on clean images, they are vulnerable to distortions, such as noise and blur. Fine-tuning a pre-trained CNN on mutually exclusive or a union set of distortions is a brute-force solution. This iterative fine-tuning process with all known types of distortion is, however, exhaustive and the network struggles to handle unseen distortions. CNNs are also vulnerable to image translation or shift, partly due to common Down-Sampling (DS) layers, e.g., max-pooling and strided convolution. These operations violate the Nyquist sampling rate and cause aliasing. The textbook solution is low-pass filtering (blurring) before down-sampling, which can benefit deep networks as well. Even so, non-linearity units, such as ReLU, often re-introduce the problem, suggesting that blurring alone may not suffice. Another important but under-explored issue for CNNs is unknown or Open Set Recognition (OSR). CNNs are commonly designed for closed set arrangements, where test instances only belong to some ‘Known Known’ (KK) classes used in training. As such, they predict a class label for a test sample based on the distribution of the KK classes. However, when used under the OSR setup (where an input may belong to an ‘Unknown Unknown’ or UU class), such a network will always classify a test instance as one of the KK classes even if it is from a UU class. Historically, CNNs have struggled with detecting objects in images with large difference in scale, especially small objects. This is because the DS layers inside a CNN often progressively wipe out the signal from small objects. As a result, the final layers are left with no signature from these objects leading to degraded performance. In this work, we propose solutions to the above four problems. First, we improve CNN robustness against distortion by proposing DCT based augmentation, adaptive regularisation, and noise suppressing Activation Functions (AF). Second, to ensure further performance gain and robustness to image transformations, we introduce anti-aliasing properties inside the AF and propose a novel DS method called blurpool. Third, to address the OSR problem, we propose a novel training paradigm that ensures detection of UU classes and accurate classification of the KK classes. Finally, we introduce a novel CNN that enables a deep detector to identify small objects with high precision and recall. We evaluate our methods on a number of benchmark datasets and demonstrate that they outperform contemporary methods in the respective problem set-ups.
- Description: Doctor of Philosophy
- Authors: Hossain, Md Tahmid
- Date: 2021
- Type: Text , Thesis , PhD
- Full Text:
- Description: Image classification is one of the fundamental tasks in the field of computer vision. Although Artificial Neural Network (ANN) showed a lot of promise in this field, the lack of efficient computer hardware subdued its potential to a great extent. In the early 2000s, advances in hardware coupled with better network design saw the dramatic rise of Convolutional Neural Network (CNN). Deep CNNs pushed the State-of-The-Art (SOTA) in a number of vision tasks, including image classification, object detection, and segmentation. Presently, CNNs dominate these tasks. Although CNNs exhibit impressive classification performance on clean images, they are vulnerable to distortions, such as noise and blur. Fine-tuning a pre-trained CNN on mutually exclusive or a union set of distortions is a brute-force solution. This iterative fine-tuning process with all known types of distortion is, however, exhaustive and the network struggles to handle unseen distortions. CNNs are also vulnerable to image translation or shift, partly due to common Down-Sampling (DS) layers, e.g., max-pooling and strided convolution. These operations violate the Nyquist sampling rate and cause aliasing. The textbook solution is low-pass filtering (blurring) before down-sampling, which can benefit deep networks as well. Even so, non-linearity units, such as ReLU, often re-introduce the problem, suggesting that blurring alone may not suffice. Another important but under-explored issue for CNNs is unknown or Open Set Recognition (OSR). CNNs are commonly designed for closed set arrangements, where test instances only belong to some ‘Known Known’ (KK) classes used in training. As such, they predict a class label for a test sample based on the distribution of the KK classes. However, when used under the OSR setup (where an input may belong to an ‘Unknown Unknown’ or UU class), such a network will always classify a test instance as one of the KK classes even if it is from a UU class. Historically, CNNs have struggled with detecting objects in images with large difference in scale, especially small objects. This is because the DS layers inside a CNN often progressively wipe out the signal from small objects. As a result, the final layers are left with no signature from these objects leading to degraded performance. In this work, we propose solutions to the above four problems. First, we improve CNN robustness against distortion by proposing DCT based augmentation, adaptive regularisation, and noise suppressing Activation Functions (AF). Second, to ensure further performance gain and robustness to image transformations, we introduce anti-aliasing properties inside the AF and propose a novel DS method called blurpool. Third, to address the OSR problem, we propose a novel training paradigm that ensures detection of UU classes and accurate classification of the KK classes. Finally, we introduce a novel CNN that enables a deep detector to identify small objects with high precision and recall. We evaluate our methods on a number of benchmark datasets and demonstrate that they outperform contemporary methods in the respective problem set-ups.
- Description: Doctor of Philosophy
Fraud detection for online banking for scalable and distributed data
- Authors: Haq, Ikram
- Date: 2020
- Type: Text , Thesis , PhD
- Full Text:
- Description: Online fraud causes billions of dollars in losses for banks. Therefore, online banking fraud detection is an important field of study. However, there are many challenges in conducting research in fraud detection. One of the constraints is due to unavailability of bank datasets for research or the required characteristics of the attributes of the data are not available. Numeric data usually provides better performance for machine learning algorithms. Most transaction data however have categorical, or nominal features as well. Moreover, some platforms such as Apache Spark only recognizes numeric data. So, there is a need to use techniques e.g. One-hot encoding (OHE) to transform categorical features to numerical features, however OHE has challenges including the sparseness of transformed data and that the distinct values of an attribute are not always known in advance. Efficient feature engineering can improve the algorithm’s performance but usually requires detailed domain knowledge to identify correct features. Techniques like Ripple Down Rules (RDR) are suitable for fraud detection because of their low maintenance and incremental learning features. However, high classification accuracy on mixed datasets, especially for scalable data is challenging. Evaluation of RDR on distributed platforms is also challenging as it is not available on these platforms. The thesis proposes the following solutions to these challenges: • We developed a technique Highly Correlated Rule Based Uniformly Distribution (HCRUD) to generate highly correlated rule-based uniformly-distributed synthetic data. • We developed a technique One-hot Encoded Extended Compact (OHE-EC) to transform categorical features to numeric features by compacting sparse-data even if all distinct values are unknown. • We developed a technique Feature Engineering and Compact Unified Expressions (FECUE) to improve model efficiency through feature engineering where the domain of the data is not known in advance. • A Unified Expression RDR fraud deduction technique (UE-RDR) for Big data has been proposed and evaluated on the Spark platform. Empirical tests were executed on multi-node Hadoop cluster using well-known classifiers on bank data, synthetic bank datasets and publicly available datasets from UCI repository. These evaluations demonstrated substantial improvements in terms of classification accuracy, ruleset compactness and execution speed.
- Description: Doctor of Philosophy
- Authors: Haq, Ikram
- Date: 2020
- Type: Text , Thesis , PhD
- Full Text:
- Description: Online fraud causes billions of dollars in losses for banks. Therefore, online banking fraud detection is an important field of study. However, there are many challenges in conducting research in fraud detection. One of the constraints is due to unavailability of bank datasets for research or the required characteristics of the attributes of the data are not available. Numeric data usually provides better performance for machine learning algorithms. Most transaction data however have categorical, or nominal features as well. Moreover, some platforms such as Apache Spark only recognizes numeric data. So, there is a need to use techniques e.g. One-hot encoding (OHE) to transform categorical features to numerical features, however OHE has challenges including the sparseness of transformed data and that the distinct values of an attribute are not always known in advance. Efficient feature engineering can improve the algorithm’s performance but usually requires detailed domain knowledge to identify correct features. Techniques like Ripple Down Rules (RDR) are suitable for fraud detection because of their low maintenance and incremental learning features. However, high classification accuracy on mixed datasets, especially for scalable data is challenging. Evaluation of RDR on distributed platforms is also challenging as it is not available on these platforms. The thesis proposes the following solutions to these challenges: • We developed a technique Highly Correlated Rule Based Uniformly Distribution (HCRUD) to generate highly correlated rule-based uniformly-distributed synthetic data. • We developed a technique One-hot Encoded Extended Compact (OHE-EC) to transform categorical features to numeric features by compacting sparse-data even if all distinct values are unknown. • We developed a technique Feature Engineering and Compact Unified Expressions (FECUE) to improve model efficiency through feature engineering where the domain of the data is not known in advance. • A Unified Expression RDR fraud deduction technique (UE-RDR) for Big data has been proposed and evaluated on the Spark platform. Empirical tests were executed on multi-node Hadoop cluster using well-known classifiers on bank data, synthetic bank datasets and publicly available datasets from UCI repository. These evaluations demonstrated substantial improvements in terms of classification accuracy, ruleset compactness and execution speed.
- Description: Doctor of Philosophy
Imbalanced data classification and its application in cyber security
- Authors: Moniruzzaman, Md
- Date: 2020
- Type: Text , Thesis , PhD
- Full Text:
- Description: Cyber security, also known as information technology security or simply as information security, aims to protect government organizations, companies and individuals by defending their computers, servers, electronic systems, networks, and data from malicious attacks. With the advancement of client-side on the fly web content generation techniques, it becomes easier for attackers to modify the content of a website dynamically and gain access to valuable information. The impact of cybercrime to the global economy is now more than ever, and it is growing day by day. Among various types of cybercrimes, financial attacks are widely spread and the financial sector is among most targeted. Both corporations and individuals are losing a huge amount of money each year. The majority portion of financial attacks is carried out by banking malware and web-based attacks. The end users are not always skilled enough to differentiate between injected content and actual contents of a webpage. Designing a real-time security system for ensuring a safe browsing experience is a challenging task. Some of the existing solutions are designed for client side and all the users have to install it in their system, which is very difficult to implement. In addition, various platforms and tools are used by organizations and individuals, therefore, different solutions are needed to be designed. The existing server-side solution often focuses on sanitizing and filtering the inputs. It will fail to detect obfuscated and hidden scripts. This is a realtime security system and any significant delay will hamper user experience. Therefore, finding the most optimized and efficient solution is very important. To ensure an easy installation and integration capabilities of any solution with the existing system is also a critical factor to consider. If the solution is efficient but difficult to integrate, then it may not be a feasible solution for practical use. Unsupervised and supervised data classification techniques have been widely applied to design algorithms for solving cyber security problems. The performance of these algorithms varies depending on types of cyber security problems and size of datasets. To date, existing algorithms do not achieve high accuracy in detecting malware activities. Datasets in cyber security and, especially those from financial sectors, are predominantly imbalanced datasets as the number of malware activities is significantly less than the number of normal activities. This means that classifiers for imbalanced datasets can be used to develop supervised data classification algorithms to detect malware activities. Development of classifiers for imbalanced data sets has been subject of research over the last decade. Most of these classifiers are based on oversampling and undersampling techniques and are not efficient in many situations as such techniques are applied globally. In this thesis, we develop two new algorithms for solving supervised data classification problems in imbalanced datasets and then apply them to solve malware detection problems. The first algorithm is designed using the piecewise linear classifiers by formulating this problem as an optimization problem and by applying the penalty function method. More specifically, we add more penalty to the objective function for misclassified points from minority classes. The second method is based on the combination of the supervised and unsupervised (clustering) algorithms. Such an approach allows one to identify areas in the input space where minority classes are located and to apply local oversampling or undersampling. This approach leads to the design of more efficient and accurate classifiers. The proposed algorithms are tested using real-world datasets. Results clearly demonstrate superiority of newly introduced algorithms. Then we apply these algorithms to design classifiers to detect malwares.
- Description: Doctor of Philosophy
- Authors: Moniruzzaman, Md
- Date: 2020
- Type: Text , Thesis , PhD
- Full Text:
- Description: Cyber security, also known as information technology security or simply as information security, aims to protect government organizations, companies and individuals by defending their computers, servers, electronic systems, networks, and data from malicious attacks. With the advancement of client-side on the fly web content generation techniques, it becomes easier for attackers to modify the content of a website dynamically and gain access to valuable information. The impact of cybercrime to the global economy is now more than ever, and it is growing day by day. Among various types of cybercrimes, financial attacks are widely spread and the financial sector is among most targeted. Both corporations and individuals are losing a huge amount of money each year. The majority portion of financial attacks is carried out by banking malware and web-based attacks. The end users are not always skilled enough to differentiate between injected content and actual contents of a webpage. Designing a real-time security system for ensuring a safe browsing experience is a challenging task. Some of the existing solutions are designed for client side and all the users have to install it in their system, which is very difficult to implement. In addition, various platforms and tools are used by organizations and individuals, therefore, different solutions are needed to be designed. The existing server-side solution often focuses on sanitizing and filtering the inputs. It will fail to detect obfuscated and hidden scripts. This is a realtime security system and any significant delay will hamper user experience. Therefore, finding the most optimized and efficient solution is very important. To ensure an easy installation and integration capabilities of any solution with the existing system is also a critical factor to consider. If the solution is efficient but difficult to integrate, then it may not be a feasible solution for practical use. Unsupervised and supervised data classification techniques have been widely applied to design algorithms for solving cyber security problems. The performance of these algorithms varies depending on types of cyber security problems and size of datasets. To date, existing algorithms do not achieve high accuracy in detecting malware activities. Datasets in cyber security and, especially those from financial sectors, are predominantly imbalanced datasets as the number of malware activities is significantly less than the number of normal activities. This means that classifiers for imbalanced datasets can be used to develop supervised data classification algorithms to detect malware activities. Development of classifiers for imbalanced data sets has been subject of research over the last decade. Most of these classifiers are based on oversampling and undersampling techniques and are not efficient in many situations as such techniques are applied globally. In this thesis, we develop two new algorithms for solving supervised data classification problems in imbalanced datasets and then apply them to solve malware detection problems. The first algorithm is designed using the piecewise linear classifiers by formulating this problem as an optimization problem and by applying the penalty function method. More specifically, we add more penalty to the objective function for misclassified points from minority classes. The second method is based on the combination of the supervised and unsupervised (clustering) algorithms. Such an approach allows one to identify areas in the input space where minority classes are located and to apply local oversampling or undersampling. This approach leads to the design of more efficient and accurate classifiers. The proposed algorithms are tested using real-world datasets. Results clearly demonstrate superiority of newly introduced algorithms. Then we apply these algorithms to design classifiers to detect malwares.
- Description: Doctor of Philosophy
Effective and efficient kernel-based image representations for classification and retrieval
- Authors: Karmakar, Priyabrata
- Date: 2018
- Type: Text , Thesis , PhD
- Full Text:
- Description: Image representation is a challenging task. In particular, in order to obtain better performances in different image processing applications such as video surveillance, autonomous driving, crime scene detection and automatic inspection, effective and efficient image representation is a fundamental need. The performance of these applications usually depends on how accurately images are classified into their corresponding groups or how precisely relevant images are retrieved from a database based on a query. Accuracy in image classification and precision in image retrieval depend on the effectiveness of image representation. Existing image representation methods have some limitations. For example, spatial pyramid matching, which is a popular method incorporating spatial information in image-level representation, has not been fully studied to date. In addition, the strengths of pyramid match kernel and spatial pyramid matching are not combined for better image matching. Kernel descriptors based on gradient, colour and shape overcome the limitations of histogram-based descriptors, but suffer from information loss, noise effects and high computational complexity. Furthermore, the combined performance of kernel descriptors has limitations related to computational complexity, higher dimensionality and lower effectiveness. Moreover, the potential of a global texture descriptor which is based on human visual perception has not been fully explored to date. Therefore, in this research project, kernel-based effective and efficient image representation methods are proposed to address the above limitations. An enhancement is made to spatial pyramid matching in terms of improved rotation invariance. This is done by investigating different partitioning schemes suitable to achieve rotation-invariant image representation and the proposal of a weight function for appropriate level contribution in image matching. In addition, the strengths of pyramid match kernel and spatial pyramid are combined to enhance matching accuracy between images. The existing kernel descriptors are modified and improved to achieve greater effectiveness, minimum noise effects, less dimensionality and lower computational complexity. A novel fusion approach is also proposed to combine the information related to all pixel attributes, before the descriptor extraction stage. Existing kernel descriptors are based only on gradient, colour and shape information. In this research project, a texture-based kernel descriptor is proposed by modifying an existing popular global texture descriptor. Finally, all the contributions are evaluated in an integrated system. The performances of the proposed methods are qualitatively and quantitatively evaluated on two to four different publicly available image databases. The experimental results show that the proposed methods are more effective and efficient in image representation than existing benchmark methods.
- Description: Doctor of Philosophy
- Authors: Karmakar, Priyabrata
- Date: 2018
- Type: Text , Thesis , PhD
- Full Text:
- Description: Image representation is a challenging task. In particular, in order to obtain better performances in different image processing applications such as video surveillance, autonomous driving, crime scene detection and automatic inspection, effective and efficient image representation is a fundamental need. The performance of these applications usually depends on how accurately images are classified into their corresponding groups or how precisely relevant images are retrieved from a database based on a query. Accuracy in image classification and precision in image retrieval depend on the effectiveness of image representation. Existing image representation methods have some limitations. For example, spatial pyramid matching, which is a popular method incorporating spatial information in image-level representation, has not been fully studied to date. In addition, the strengths of pyramid match kernel and spatial pyramid matching are not combined for better image matching. Kernel descriptors based on gradient, colour and shape overcome the limitations of histogram-based descriptors, but suffer from information loss, noise effects and high computational complexity. Furthermore, the combined performance of kernel descriptors has limitations related to computational complexity, higher dimensionality and lower effectiveness. Moreover, the potential of a global texture descriptor which is based on human visual perception has not been fully explored to date. Therefore, in this research project, kernel-based effective and efficient image representation methods are proposed to address the above limitations. An enhancement is made to spatial pyramid matching in terms of improved rotation invariance. This is done by investigating different partitioning schemes suitable to achieve rotation-invariant image representation and the proposal of a weight function for appropriate level contribution in image matching. In addition, the strengths of pyramid match kernel and spatial pyramid are combined to enhance matching accuracy between images. The existing kernel descriptors are modified and improved to achieve greater effectiveness, minimum noise effects, less dimensionality and lower computational complexity. A novel fusion approach is also proposed to combine the information related to all pixel attributes, before the descriptor extraction stage. Existing kernel descriptors are based only on gradient, colour and shape information. In this research project, a texture-based kernel descriptor is proposed by modifying an existing popular global texture descriptor. Finally, all the contributions are evaluated in an integrated system. The performances of the proposed methods are qualitatively and quantitatively evaluated on two to four different publicly available image databases. The experimental results show that the proposed methods are more effective and efficient in image representation than existing benchmark methods.
- Description: Doctor of Philosophy
Pixel N-grams for Mammographic Image Classification
- Authors: Kulkarni, Pradnya
- Date: 2017
- Type: Text , Thesis , PhD
- Full Text:
- Description: X-ray screening for breast cancer is an important public health initiative in the management of a leading cause of death for women. However, screening is expensive if mammograms are required to be manually assessed by radiologists. Moreover, manual screening is subject to perception and interpretation errors. Computer aided detection/diagnosis (CAD) systems can help radiologists as computer algorithms are good at performing image analysis consistently and repetitively. However, image features that enhance CAD classification accuracies are necessary for CAD systems to be deployed. Many CAD systems have been developed but the specificity and sensitivity is not high; in part because of challenges inherent in identifying effective features to be initially extracted from raw images. Existing feature extraction techniques can be grouped under three main approaches; statistical, spectral and structural. Statistical and spectral techniques provide global image features but often fail to distinguish between local pattern variations within an image. On the other hand, structural approach have given rise to the Bag-of-Visual-Words (BoVW) model, which captures local variations in an image, but typically do not consider spatial relationships between the visual “words”. Moreover, statistical features and features based on BoVW models are computationally very expensive. Similarly, structural feature computation methods other than BoVW are also computationally expensive and strongly dependent upon algorithms that can segment an image to localize a region of interest likely to contain the tumour. Thus, classification algorithms using structural features require high resource computers. In order for a radiologist to classify the lesions on low resource computers such as Ipads, Tablets, and Mobile phones, in a remote location, it is necessary to develop computationally inexpensive classification algorithms. Therefore, the overarching aim of this research is to discover a feature extraction/image representation model which can be used to classify mammographic lesions with high accuracy, sensitivity and specificity along with low computational cost. For this purpose a novel feature extraction technique called ‘Pixel N-grams’ is proposed. The Pixel N-grams approach is inspired from the character N-gram concept in text categorization. Here, N number of consecutive pixel intensities are considered in a particular direction. The image is then represented with the help of histogram of occurrences of the Pixel N-grams in an image. Shape and texture of mammographic lesions play an important role in determining the malignancy of the lesion. It was hypothesized that the Pixel N-grams would be able to distinguish between various textures and shapes. Experiments carried out on benchmark texture databases and binary basic shapes database have demonstrated that the hypothesis was correct. Moreover, the Pixel N-grams were able to distinguish between various shapes irrespective of size and location of shape in an image. The efficacy of the Pixel N-gram technique was tested on mammographic database of primary digital mammograms sourced from a radiological facility in Australia (LakeImaging Pty Ltd) and secondary digital mammograms (benchmark miniMIAS database). A senior radiologist from LakeImaging provided real time de-identified high resolution mammogram images with annotated regions of interests (which were used as groundtruth), and valuable radiological diagnostic knowledge. Two types of classifications were observed on these two datasets. Normal/abnormal classification useful for automated screening and circumscribed/speculation/normal classification useful for automated diagnosis of breast cancer. The classification results on both the mammography datasets using Pixel N-grams were promising. Classification performance (Fscore, sensitivity and specificity) using Pixel N-gram technique was observed to be significantly better than the existing techniques such as intensity histogram, co-occurrence matrix based features and comparable with the BoVW features. Further, Pixel N-gram features are found to be computationally less complex than the co-occurrence matrix based features as well as BoVW features paving the way for mammogram classification on low resource computers. Although, the Pixel N-gram technique was designed for mammographic classification, it could be applied to other image classification applications such as diabetic retinopathy, histopathological image classification, lung tumour detection using CT images, brain tumour detection using MRI images, wound image classification and tooth decay classification using dentistry x-ray images. Further, texture and shape classification is also useful for classification of real world images outside the medical domain. Therefore, the pixel N-gram technique could be extended for applications such as classification of satellite imagery and other object detection tasks.
- Description: Doctor of Philosophy
- Authors: Kulkarni, Pradnya
- Date: 2017
- Type: Text , Thesis , PhD
- Full Text:
- Description: X-ray screening for breast cancer is an important public health initiative in the management of a leading cause of death for women. However, screening is expensive if mammograms are required to be manually assessed by radiologists. Moreover, manual screening is subject to perception and interpretation errors. Computer aided detection/diagnosis (CAD) systems can help radiologists as computer algorithms are good at performing image analysis consistently and repetitively. However, image features that enhance CAD classification accuracies are necessary for CAD systems to be deployed. Many CAD systems have been developed but the specificity and sensitivity is not high; in part because of challenges inherent in identifying effective features to be initially extracted from raw images. Existing feature extraction techniques can be grouped under three main approaches; statistical, spectral and structural. Statistical and spectral techniques provide global image features but often fail to distinguish between local pattern variations within an image. On the other hand, structural approach have given rise to the Bag-of-Visual-Words (BoVW) model, which captures local variations in an image, but typically do not consider spatial relationships between the visual “words”. Moreover, statistical features and features based on BoVW models are computationally very expensive. Similarly, structural feature computation methods other than BoVW are also computationally expensive and strongly dependent upon algorithms that can segment an image to localize a region of interest likely to contain the tumour. Thus, classification algorithms using structural features require high resource computers. In order for a radiologist to classify the lesions on low resource computers such as Ipads, Tablets, and Mobile phones, in a remote location, it is necessary to develop computationally inexpensive classification algorithms. Therefore, the overarching aim of this research is to discover a feature extraction/image representation model which can be used to classify mammographic lesions with high accuracy, sensitivity and specificity along with low computational cost. For this purpose a novel feature extraction technique called ‘Pixel N-grams’ is proposed. The Pixel N-grams approach is inspired from the character N-gram concept in text categorization. Here, N number of consecutive pixel intensities are considered in a particular direction. The image is then represented with the help of histogram of occurrences of the Pixel N-grams in an image. Shape and texture of mammographic lesions play an important role in determining the malignancy of the lesion. It was hypothesized that the Pixel N-grams would be able to distinguish between various textures and shapes. Experiments carried out on benchmark texture databases and binary basic shapes database have demonstrated that the hypothesis was correct. Moreover, the Pixel N-grams were able to distinguish between various shapes irrespective of size and location of shape in an image. The efficacy of the Pixel N-gram technique was tested on mammographic database of primary digital mammograms sourced from a radiological facility in Australia (LakeImaging Pty Ltd) and secondary digital mammograms (benchmark miniMIAS database). A senior radiologist from LakeImaging provided real time de-identified high resolution mammogram images with annotated regions of interests (which were used as groundtruth), and valuable radiological diagnostic knowledge. Two types of classifications were observed on these two datasets. Normal/abnormal classification useful for automated screening and circumscribed/speculation/normal classification useful for automated diagnosis of breast cancer. The classification results on both the mammography datasets using Pixel N-grams were promising. Classification performance (Fscore, sensitivity and specificity) using Pixel N-gram technique was observed to be significantly better than the existing techniques such as intensity histogram, co-occurrence matrix based features and comparable with the BoVW features. Further, Pixel N-gram features are found to be computationally less complex than the co-occurrence matrix based features as well as BoVW features paving the way for mammogram classification on low resource computers. Although, the Pixel N-gram technique was designed for mammographic classification, it could be applied to other image classification applications such as diabetic retinopathy, histopathological image classification, lung tumour detection using CT images, brain tumour detection using MRI images, wound image classification and tooth decay classification using dentistry x-ray images. Further, texture and shape classification is also useful for classification of real world images outside the medical domain. Therefore, the pixel N-gram technique could be extended for applications such as classification of satellite imagery and other object detection tasks.
- Description: Doctor of Philosophy
The effectiveness of using static features in identifying scam genres
- Authors: Stabek, Amber
- Date: 2010
- Type: Text , Thesis , Masters
- Full Text:
- Description: Thesis details a cybercrime classification framework stemming from a mixed methodological approach, which is both top down and bottom up and is designed to be multidisciplinary and adaptable across sectors.
- Description: Master by Research of Mathematical Sciences
- Description: Variation in scam classification is regularly identified as a primary cause of discrepancy in victim report data resulting in unsuccessful scam identification and insufficient rates of interception by law enforcement, which results in the low prosecution rate of scammers. The result of such discrepancies lead to complex concerns, such as the under reporting of scam incidence, and reduced rates of successful follow up by investigative and enforcement agencies consequential to difficulties in making correct referrals. Without a shared and common lexicon of scam labels and descriptions, communication between investigative agencies and cross-border cooperation is obstructed. With no compatible comprehension of the scam lexicon, timely progression in scam-case management leading to the identification, tracking and interception of scammer communications cannot be realised. Ambiguities leading to interpretational impedances are aiding scammers by enabling their scams in cross-jurisdictional and multi-national platforms. If the wide variety of known scam types could be condensed to recognisable and traceable instances, the business models that scammers use could be identified and future scamming events predicted, monitored, and interrupted. Following a mixed methodology, this research aims to address some of these concerns. This is achieved by clustering scam descriptions and partitioning them into scam types, called scam genres. The result of which reveals homogeneous groups of scam cases and allows for the assessment of the effectiveness of using static features in identifying scam types. Second to this, identification of the most suitable model for reducing scam cases into the fewest number of clusters with the least number of scam cases within in each cluster at an accuracy level of at least 95% is achieved. Through the use of hierarchical clustering, this research grouped publically available scams into homogeneous clusters of scam genres. Two-hundred and seventy-seven scams from 38 separate categories of scam classification were condensed into as few as 7-clusters of scam genre. Following a mixed methodological, grounded theoretical approach and using discriminant function analysis, 82 static features were derived from the 277 scam descriptions analysed. Of the 82 static features derived, it was concluded that only 68 significantly predicted scam type and explained 95% of the total variation found in scam case assignment. The most significant static features determined to be crucial to any scamming campaign and useful in identifying the type of scam genre a scam case belongs to were; what the scam offered, the role of the victim, the goal of the scammer and the method of scam introduction. The results of this research provide empirical evidence of the inconsistent use of definitions across jurisdictions in scam descriptions, and will contribute to the development of a uniform lexicon of scamming terminology as well as become foundational to further research on the impact of scams for law enforcement, the public and private sector, the community and the individual.
- Authors: Stabek, Amber
- Date: 2010
- Type: Text , Thesis , Masters
- Full Text:
- Description: Thesis details a cybercrime classification framework stemming from a mixed methodological approach, which is both top down and bottom up and is designed to be multidisciplinary and adaptable across sectors.
- Description: Master by Research of Mathematical Sciences
- Description: Variation in scam classification is regularly identified as a primary cause of discrepancy in victim report data resulting in unsuccessful scam identification and insufficient rates of interception by law enforcement, which results in the low prosecution rate of scammers. The result of such discrepancies lead to complex concerns, such as the under reporting of scam incidence, and reduced rates of successful follow up by investigative and enforcement agencies consequential to difficulties in making correct referrals. Without a shared and common lexicon of scam labels and descriptions, communication between investigative agencies and cross-border cooperation is obstructed. With no compatible comprehension of the scam lexicon, timely progression in scam-case management leading to the identification, tracking and interception of scammer communications cannot be realised. Ambiguities leading to interpretational impedances are aiding scammers by enabling their scams in cross-jurisdictional and multi-national platforms. If the wide variety of known scam types could be condensed to recognisable and traceable instances, the business models that scammers use could be identified and future scamming events predicted, monitored, and interrupted. Following a mixed methodology, this research aims to address some of these concerns. This is achieved by clustering scam descriptions and partitioning them into scam types, called scam genres. The result of which reveals homogeneous groups of scam cases and allows for the assessment of the effectiveness of using static features in identifying scam types. Second to this, identification of the most suitable model for reducing scam cases into the fewest number of clusters with the least number of scam cases within in each cluster at an accuracy level of at least 95% is achieved. Through the use of hierarchical clustering, this research grouped publically available scams into homogeneous clusters of scam genres. Two-hundred and seventy-seven scams from 38 separate categories of scam classification were condensed into as few as 7-clusters of scam genre. Following a mixed methodological, grounded theoretical approach and using discriminant function analysis, 82 static features were derived from the 277 scam descriptions analysed. Of the 82 static features derived, it was concluded that only 68 significantly predicted scam type and explained 95% of the total variation found in scam case assignment. The most significant static features determined to be crucial to any scamming campaign and useful in identifying the type of scam genre a scam case belongs to were; what the scam offered, the role of the victim, the goal of the scammer and the method of scam introduction. The results of this research provide empirical evidence of the inconsistent use of definitions across jurisdictions in scam descriptions, and will contribute to the development of a uniform lexicon of scamming terminology as well as become foundational to further research on the impact of scams for law enforcement, the public and private sector, the community and the individual.
Classification of HTML Documents
- Xie, Wei
- Authors: Xie, Wei
- Date: 2006
- Type: Text , Thesis , PhD
- Full Text:
- Description: Text Classification is the task of mapping a document into one or more classes based on the presence or absence of words (or features) in the document. It is intensively being studied and different classification techniques and algorithms have been developed. This thesis focuses on classification of online documents that has become more critical with the development of World Wide Web. The WWW vastly increases the availability of on-line documents in digital format and has highlighted the need to classify them. From this background, we have noted the emergence of “automatic Web Classification”. These mainly concentrate on classifying HTML-like documents into classes or categories by not only using the methods that are inherited from the traditional Text Classification process, but also utilizing the extra information provided only by Web pages. Our work is based on the fact that, Web documents, contain not only ordinary features (words) but also extra information, such as meta-data and hyperlinks that can be used to advantage the classification process. The aim of this research is to study various ways of using the extra information, in particularly, hyperlink information provided by HTML-documents (Web pages). The merit of the approach, developed in this thesis, is its simplicity, compared with existing approaches. We present different approaches of using hyperlink information to improve the effectiveness of web classification. Unlike other work in this area, we will only use the mappings between linked documents and their own class or classes. In this case, we only need to add a few features called linked-class features into the datasets, and then apply classifiers on them for classification. In the numerical experiments we adopted two wellknown Text Classification algorithms, Support Vector Machines and BoosTexter. The results obtained show that classification accuracy can be improved by using mixtures of ordinary and linked-class features. Moreover, out-links usually work better than in-links in classification. We also analyse and discuss the reasons behind this improvement.
- Description: Master of Computing
- Authors: Xie, Wei
- Date: 2006
- Type: Text , Thesis , PhD
- Full Text:
- Description: Text Classification is the task of mapping a document into one or more classes based on the presence or absence of words (or features) in the document. It is intensively being studied and different classification techniques and algorithms have been developed. This thesis focuses on classification of online documents that has become more critical with the development of World Wide Web. The WWW vastly increases the availability of on-line documents in digital format and has highlighted the need to classify them. From this background, we have noted the emergence of “automatic Web Classification”. These mainly concentrate on classifying HTML-like documents into classes or categories by not only using the methods that are inherited from the traditional Text Classification process, but also utilizing the extra information provided only by Web pages. Our work is based on the fact that, Web documents, contain not only ordinary features (words) but also extra information, such as meta-data and hyperlinks that can be used to advantage the classification process. The aim of this research is to study various ways of using the extra information, in particularly, hyperlink information provided by HTML-documents (Web pages). The merit of the approach, developed in this thesis, is its simplicity, compared with existing approaches. We present different approaches of using hyperlink information to improve the effectiveness of web classification. Unlike other work in this area, we will only use the mappings between linked documents and their own class or classes. In this case, we only need to add a few features called linked-class features into the datasets, and then apply classifiers on them for classification. In the numerical experiments we adopted two wellknown Text Classification algorithms, Support Vector Machines and BoosTexter. The results obtained show that classification accuracy can be improved by using mixtures of ordinary and linked-class features. Moreover, out-links usually work better than in-links in classification. We also analyse and discuss the reasons behind this improvement.
- Description: Master of Computing
- «
- ‹
- 1
- ›
- »