Graham Sexton

An adaptive ensemble classifier for mining concept drifting data streams

It is challenging to use traditional data mining techniques to deal with real-time data stream classifications. Existing mining classifiers need to be updated frequently to adapt to the changes in data streams. To address this issue, in this paper we propose an adaptive ensemble approach for classification and novel class detection in concept drifting data streams. The proposed approach uses traditional mining classifiers and updates the ensemble model automatically so that it represents the most recent concepts in data streams. For novel class detection we consider the idea that data points belonging to the same class should be closer to each other and should be far apart from the data points belonging to other classes. If a data point is well separated from the existing data clusters, it is identified as a novel class instance. We tested the performance of this proposed stream classification model against that of existing mining algorithms using real benchmark datasets from UCI (University of California, Irvine) machine learning repository. The experimental results prove that our approach shows great flexibility and robustness in novel class detection in concept drifting and outperforms traditional classification models in challenging real-life data stream applications.

Intelligent phishing detection and protection scheme for online transactions

Phishing is an instance of social engineering techniques used to deceive users into giving their sensitive information using an illegitimate website that looks and feels exactly like the target organization website. Most phishing detection approaches utilizes Uniform Resource Locator (URL) blacklists or phishing website features combined with machine learning techniques to combat phishing. Despite the existing approaches that utilize URL blacklists, they cannot generalize well with new phishing attacks due to human weakness in verifying blacklists, while the existing feature-based methods suffer high false positive rates and insufficient phishing features. As a result, this leads to an inadequacy in the online transactions. To solve this problem robustly, the proposed study introduces new inputs (Legitimate site rules, User-behavior profile, PhishTank, User-specific sites, Pop-Ups from emails) which were not considered previously in a single protection platform. The idea is to utilize a Neuro-Fuzzy Scheme with 5 inputs to detect phishing sites with high accuracy in real-time. In this study, 2-Fold cross-validation is applied for training and testing the proposed model. A total of 288 features with 5 inputs were used and has so far achieved the best performance as compared to all previously reported results in the field.

Speaker identification using multilayer perceptrons and radial basis function networks

Abstract This paper compares the Multilayer Perceptrons network (trained by the backpropagation algorithm) and the Radial Basis Function networks in the task of speaker identification. The experiments were carried out on 200 utterances (10 digits) of 10 speakers. LPC-derived cepstrum coefficients were used as the speaker specific features. The results showed that the Multilayer Perceptrons networks were superior in memory usage and classification time. However, they suffered from long training time and the error rate was slightly higher than that of Radial Basis Function networks.

Multi-objective multi-drug scheduling schemes for cell cycle specific cancer treatment

This paper presents an investigation into the development of an optimal chemotherapy drug(s) scheduling scheme to control the drug doses to be infused to the patients body. The current standard of practice of treatment is based on empirical evidence gathered from preclinical and clinical trials carried out during the drug development process. In general, most chemotherapy drugs used in cancer treatments are toxic agents and usually have narrow therapeutic indices; dose levels at which these drugs significantly kill the cancerous cells are close to those levels at which harmful toxic side effects occur. Therefore, an effective chemotherapy treatment protocol requires advanced automation and treatment design tools for use in clinical practice and the challenges inherent to complex biomedical systems and clinical deployment of technology (Parker, 2009). An optimum but effective drug scheduling requires suitable balancing between the beneficial and toxic side effects. Conventional clinical methods very often fail to find right drug doses that balance between these two constraints due to their inherent conflicting nature. A Multi-objective Genetic Algorithm Optimization (MOGA) process is employed to find the desired drug concentration at tumour sites that trade-off between the conflicting objectives. A close-loop control method, namely Integral-Proportional-Derivative (I-PD) is designed to control the drug to be infused to the patients body and MOGA is used to find suitable/acceptable drug concentration at tumour site and parameters of the controller. Cell cycle specific cancer tumour models have been used in this work to show the effects of drug(s) on different cell populations, drug concentrations and toxic side effects. Results show that the applied multi-objective optimization approach can produce a wide range of solutions that trade-off between cell killing and toxic side effects and satisfy associated goals of chemotherapy treatment. Depending on the physiological state of the patient and state of the cancer, the oncologist can pick the right solution suitable for the patient. The chemotherapy drug schedules obtained by the proposed treatment protocols appears to be continuous on the time (day) scale, i.e., specific amount of drugs to be administered to the patient on daily basis which can be termed as Metronomics in nature. The dose duration and the interval period between dose applications can be adjusted in the proposed scheme either by setting the sampling time of closed-loop I-PD controller to any value depending on the state of the patient and disease (model parameters) or by using genetic optimization process aiming to minimize/maximize treatment objectives and satisfying treatment constraints. Regarding the total duration of the treatment, clinical knowledge can be utilized giving emphasis on physiological state of the patient, state of the tumour and disease. Moreover, the total duration of the treatment can also be found/determined for specific values of model parameters describing physiological state of the patient, state of the tumour and disease through multi-objective optimization process. It is noted that the proposed scheme offered the best treatment performance as compared to the reported work available so far. Moreover, robustness analysis shows that the control scheme is highly stable and robust despite the model uncertainties; from small to wide range, and the percentage of proliferating cell reduction is almost same as it is found with optimum model parameters without having any uncertainty.

Intrusion detection system by fuzzy interpolation

Network intrusion detection systems identify malicious connections and thus help protect networks from attacks. Various data-driven approaches have been used in the development of network intrusion detection systems, which usually lead to either very complex systems or poor generalization ability due to the complexity of this challenge. This paper proposes a data-driven network intrusion detection system using fuzzy interpolation in an effort to address the aforementioned limitations. In particular, the developed system equipped with a sparse rule base not only guarantees the online performance of intrusion detection, but also allows the generation of security alerts from situations which are not directly covered by the existing knowledge base. The proposed system has been applied to a well-known data set for system validation and evaluation with competitive results generated.

Visual Clustering of Trademarks Using the Self-Organizing Map

This paper describes the experiments used to investigate ways in which digitised trademark images can be visually clustered on a 2-D surface, using the topological properties of the self-organizing map. Experiments were carried out on a set of original and edge detected binary trademark images, as well as their moment invariants, angular radial transformations and wavelet feature vectors. A radial based precision-recall measure was also used to evaluate the results objectively. Initial results are encouraging.

Please slow down!: the impact on tor performance from mobility

The number of mobile devices, connecting to the Internet, is predicted to surpass desktop connections by 2014. The likely growth in their mobile client base will offer an additional challenge for anonymity networks, such as Tor, in maintaining an efficient privacy service. We have conducted a simple experiment that illustrates this challenge. We have simulated the performance achieved by a mobile Tor node as it roams at varying speeds between wireless networks. The results show that the impact on performance for the mobile user, and potentially the wider Tor network, is significant when roaming, and as expected, increases with higher mobility speeds and longer recovery times. We review a range of solutions and suggest that, although the use of a lighter transport protocol and/or adaptive client throttling may reduce the performance impact of mobility, a better strategy is to provide a persistent connection to the Tor network for roaming mobile users.

Bayesian model for mobility prediction to support routing in Mobile Ad-Hoc Networks

This paper introduces a Bayesian model to predict and classify the mobility of a node in Mobile Ad-hoc Networks (MANETs). The proposed model does not use the additional information from Global Positioning System (GPS) for its prediction as some existing models did. Instead, it relies on the “average encounter rate” and “node degree” calculated at each node. However, the outcome is still recorded at high accuracy, i.e. prediction error is fewer than 10% at high speed level (above 15m/s). The aim of this model is to help a routing protocol in MANETs avoid broadcasting request messages from a high mobility node/region relied on the outcome of the prediction. Through simulation experiments, route error rate observed reduced significantly compared to normal broadcast scheme of the Ad-hoc On-demand Distance Vector (AODV) protocol. The packet delivery ratio improved up to 46.32% at the maximum velocity of 30m/s (equal to 108km/h) in the density of 200nodes/km2.

Self-adaptive proactive routing scheme for mobile ad-hoc networks

This study introduces a routing model which has the ability to detect the mobile ad-hoc network (MANET) mobility states and self-adapt routing metrics accordingly. In this model, nodes rely on theirs mobility indicator to detect whether the network is relatively static or mobile and hence switch the routing metric to either expected transmission count (ETX) or mobility factor (MF), respectively. The proposed model takes advantages of both ETX and MF metrics thus enhancing the overall routing performance for MANET in different mobility states. Packet delivery ratio increases 10% in both static and mobile conditions whereas the number of drop packets reduces half compared with the original optimised link state routing protocol.

An Extended Takagi-Sugeno-Kang Inference System (TSK+) with Fuzzy Interpolation and Its Rule Base Generation

A rule base covering the entire input domain is required for the conventional Mamdani inference and Takagi–Sugeno–Kang (TSK) inference. Fuzzy interpolation enhances conventional fuzzy rule inference systems by allowing the use of sparse rule bases by which certain inputs are not covered. Given that almost all of the existing fuzzy interpolation approaches were developed to support the Mamdani inference, this paper presents a novel fuzzy interpolation approach that extends the TSK inference. This paper also proposes a data-driven rule base generation method to support the extended TSK inference system. The proposed system enhances the conventional TSK inference in two ways: (1) workable with incomplete or unevenly distributed data sets or incomplete expert knowledge that entails only a sparse rule base and (2) simplifying complex fuzzy inference systems by using more compact rule bases for complex systems without the sacrificing of system performance. The experimentation shows that the proposed system overall outperforms the existing approaches with the utilisation of smaller rule bases.