Browsing by Author "Nyathi, Thambo"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    Beacon Frame Manipulation to Mitigate Rogue Access Points: Case of Android Smartphone Rogue Access Points
    (Compusoft, 2014-02) Nyathi, Thambo; Ndlovu, Siqabukile
    The use of wireless devices to access corporate network resources is now part of the norm within corporate environments. When wireless users need to connect to a network they hardly question the source of their connectivity. Mobile phones, particularly smartphones allow users to access network resources. These harmless looking wireless devices can be a source of major threats if configured to be so. The Internet is awash with mobile apps capable of performing packet sniffing. These applications, coupled with the capability of the smartphone to be configured as an access point, can present a Smartphone Rogue Access Point. Access Points advertise their availability using what is called a beacon frame. This research paper proposes a solution which restructures this beacon frame to include an Authentic Access Point Value which can be used to defend against Rogue Access Points.
  • Thumbnail Image
    Item
    Bio-metric Implementation of Bayesian Networks for Face Recognitionon Android Mobile Devices
    (2013-02-01) Masundire, Daniel; Chilumani, Khesani R.; Nyathi, Thambo
    Face recognition is a field that receives attention and inquiry by Artificial Intelligence, Computer Security and Intelligent Data Processing researchers. Its application as a biometric has also grown due to the growth of the camera technologies. The digital camera’s resolution and noise reduction in images makes the camera view the world in more or less the same as a human being. However, this technology has been difficult to implement on mobile devices, phones in particular. This has been chiefly because of the computational complexity of the algorithms used. The memory size and processing speed of the mobile phones has been a constraint as well. There has been a huge growth experienced in the mobile phones’ industry with respect to the gadgets’ processing power. Face detection has been implemented using other statistical methods but this research uses the Bayesian network. This method is graphical and uses probabilistic inference. This helps in reasoning with incomplete and unknown information. The Bayesian Network is applied on data that would have acquired from a face image using Principal Component Analysis (PCA). PCA is used to reduce the dimensions of the image. The co-variances matrix from the principal components is used in the Bayesian network. The Bayesian network uses prior knowledge and the weights on the nodes are modified using weights formulated by a probability distribution function. This means, unlike other neural networks, the weights of a Bayesian network are not constants. We formulate our problem as the maximum a posteriori (MAP) estimate of a properly defined probability distribution function (PDF). A Bayesian network is used to represent the PDF as well as the domain knowledge needed for interpretation (Kumar and Desai, 1996). The development is done for Android-based mobile phones. The Android is an open source mobile phone operating system that runs on a Linux kernel. In this dissertation, we seek to improve face recognition for mobile devices by using dedicated statistical generative models. Gaussian Mixture Models (GMM) and Hidden Markov Models (HMM), which are classical generative models, have been successfully proven in face recognition. The GMM and HMM based models treat the facial features under investigation as independent. This is not consistent for a face because facial features are related to each other. Bayesian networks can be used to analyse and process the face as a graphical representation of dependent features. It only detects the frontal view of the face, that is, the face to be detected has to be showing the frontal features of the face. An appearance-based recognition and detection method is used. The prototype was built for Androidbased mobile phones. Mobile phones which run on Symbian, Windows and Apple (iOS) operating systems are not going to be covered.
  • Thumbnail Image
    Item
    Optimisation of the Linear Probability Model for Credit Risk Management
    (2014-11) Nyathi, K.T.; Ndlovu, Siqabukile; Moyo, S.; Nyathi, Thambo
    One of the aims of the banking business is to provide loans to applicants. Credit risk management plays an important role in banks, as loans generally account for half to three quarters of the total value of bank assets. Credit scoring is a systematic method for evaluating credit risk and assists decision makers determine whether or not to provide loans to applicants. Scoring models are systematic means of evaluating the creditworthiness of a loan applicant. However, existing scoring models cause some loan applications to be rejected unnecessarily as their credit rates are lowered to rejection levels due to lack of information such as previous loan payment data. This might be refusal of good credit, which potentially can cause the loss of future profit margins. This study aims at optimising one such credit scoring model to ensure that it uses only the critical scoring criteria to determine a credit score. The optimised model will not only reduce the proportion of unsafe borrowers, but also identify profitable borrowers.
  • Thumbnail Image
    Item
    Using Fuzzy ARTMAP for Symmetric Key Generation
    (American Institute of Science., 2015) Mulopa, John; Ndlovu, Siqabukile; Mzelikahle, Kernan; Nyathi, Thambo
    Neural cryptography deals with the problem of key exchange between two communicating neural networks using the mutual learning concept. It is the first algorithm for key generation over public channels which are not based on the number theory. The two networks exchange their outputs and the key between the two communicating parties is eventually presented in the final learned weights, when the two networks are synchronised. The security of neural synchronisation is put at risk if an attacker is capable of synchronising with any of the two parties during the training processes. However, the security of a cryptosystem is robust if the algorithm is strong and the keys are long, unpredictable, and random This research proposes use of two distant remote Adaptive Resonance Theory MAP (ARTMAP) architectures that are trained to learn from a unique data set and finally synchronise to same weights.