Jyothi Vadisala, Awarded

Social networks are being applied to a wide variety of applications. Social networks model the social activities between individuals, that change as time goes by. Data privacy is a major problem that has to be considered for dynamic networks before releasing network data to the public or third parties like researchers and analyzers that would compute statistics or make a deep analysis of these data. As the demand for dynamic social network analysis increases, the privacy issue in sequential publishing of social network needs to be considered. In this context, many different anonymization techniques have been proposed in the literature.

This thesis aims to address the privacy risks in dynamic social networks and to contribute a few techniques for data anonymization while optimising the privacy- utility tradeoff. Specifically, the proposed models deal with emerging privacy issues in sequential publishing of social network data. These type of data are derived from activities of our everyday life, thus the privacy of such data is impor- tant. The main objective of this thesis work is divided into two research problems.

First research problem, addresses the privacy risks of identity disclosures in sequential releases of a dynamic network. The aim is to propose a privacy model that provides adequate privacy protection and retains enough graph data utility in social network data publication. Specifically, this proposed kw-NMF anonymity model, protects identity disclosure in the context of number of common friends as a background knowledge of an adversary. To achieve this kw-NMF anonymity Number of Mutual Friend Anonymization (NMFA) greedy algorithm is proposed for large scale graph data by adding edges in anonymization process.

Performance evaluation is done on both real and synthetic data sets. The results show that the dynamic kw-NMF anonymity model can retain much of the characteristics of the network while confirming the privacy protection. The metrics used for evaluation are Graph Modification, Average Shortest Path Lengths and Average Clustering Coefficients from the aspects of protections against different information hysteresis (w) and different protection level requirements (k). Experimental results shows that the proposed algorithm can retain much of the characteristics of a dynamic network while confirming the privacy protection.

Second research problem, focuses on how the anonymization methods preserves in the existing communities of the original social networks. The k-degree and k-NMF anonmyization methods are used for anonymizing original social network and de- tect the communities for both original and anymomized networks by heuristic algorithm modular optimization Louvain method. The preservation of community is measured by Percentage of Naive Community Preservation (%NCP) and Percentage of Community Preservation at Node Level (%CPNL).

The performances of the two anonymized social networks are compared with the original social network. The experimental results show that the k-NMF anonymization method preserves the most of the communities of the original social network than k-degree anonymization method.

D Radha, Awarded

The thesis has proposed  a few techniques for data anonymization while preserving the privacy for numerical and categorical sensitive attribute based on Bucketisation.

B. Mamatha

Ocean surveillance is becoming significant since the military coastal activities are becoming more and more frequent, the ship types are increasing fast and the unconventional activities happen often. The traditional sea surveillance methods of monitoring the sea targets only by the metric information such as range, aspect, speed and direction of the target is do not giving proper results for identification of a ship. So, there is a requirement to develop automatic ship recognition systems. A few such recognition systems with ship detection software already exist. Now the challenge is to quantify detection capability and to design new software systems to better detect ship signatures to classify the ship targets of interest. In this work an Feature selection and extraction of Inverse Synthetic Aperture Radar (ISAR) image for ship classification is taken as a research problem. Ship ISAR image is a high resolution radar image.  Radar is a popular sensor for it can be used to acquire target signatures in all conditions and also from hundreds of kilometers of distance. The primary objective is to acquire target characteristics or features from the target signatures acquired through radar. Then these features are used for recognition and classification of targets. ISAR images are radar images acquired through high resolution radar which are widely being used for recognition of sea, air, land and even human targets since they contain useful information of target geometry.

In this work a set of ship ISAR images measured with high resolution radar are taken as a data set. The sea clutter and noise from nearby terrain causes hindrance to obtain clear ship ISAR images. Hence preprocessing of ISAR images is the first step and extraction of ship features having highly discriminative nature and low dimensionality from noise free ISAR images is the next step in ship classification. The problem of classification can be solved by using neural networks. Different feature extraction techniques are applied to the preprocessed ISAR images from the data sets.  The features computed from training data set are used to train the classifier and the features obtained from test set are utilized to quantify the classification accuracy of the feature vector. In this work probabilistic neural network is used as a classifier. In this work study is carried out to identify feature vectors that classify the unique set of ship ISAR images considered in this work. Various statistical techniques and mathematical transforms are used to derive features from ship ISAR images. The statistical descriptors and Zernike moments are found to give good classification accuracy. 

ISAR images are wave decomposed up to five levels with the application of multi resolution wavelet transform. For each wave, decomposed level Image, all approximation and detailed wave coefficients are computed and taken as feature vectors to identify the best suitable wave coefficients for the unique set of ship ISAR images considered in this work. Segmentation is used to find regions of interest in case of medical images, border or shape information of ISAR images or to identify linked homogeneous regions in high resolution remote sensing image analysis applications. Hence keeping in view of the significant role of segmentation in image analysis, above mentioned wave decomposed image features are also studied to find the classification accuracy of the ISAR images with segmentation.

Ship ISAR images are color images. The moments computed from individual color component images R,G,B and combined color component images  RG,GB,RB of the ISAR images taken together are considered as feature vector. Using the multi resolution wavelet transform, the ISAR images are decomposed up to five levels. The average wave energy values are computed for each of the wave decomposed images. Obtained wave energy values  are found to form a good feature vector for the ship ISAR image classification. 

In classification problems, the feature vector plays a significant role since the computational complexity and accuracy of classification depend on the features that constitute a feature vector and also the number of features in the feature vector.  All the features in a feature vector under consideration may not equally contribute to the classification of the images. Some of them may be redundant. Since a feature vector of smaller size with high discriminating nature is always preferred. Hence there is always a need to select an optimum feature set to solve the classification problem at hand. Several available optimization techniques can be explored to select optimum feature set. In this research work, ship ISAR image classification problem is studied as a single objective optimization problem and the different optimum feature combinations that give the satisfactory classification for the considered data set are identified. The optimization techniques like genetic algorithm and particle swarm optimization techniques are employed to identify the optimum feature sets in case of color moments and wave energy levels.

Samer Hasan, January 2019 Awarded

Cloud computing is a paradigm for delivering ubiquitous and on demand resources based on pay as you use financial model. It turns the IT services into utility like: water, electricity, gas and telephony. Cloud service discovery and selection becomes a significant challenge because of exponential growth in the number of cloud service providers. Typically, Cloud service providers publish cloud service advertisements and Service Level Agreement (SLA) details in various formats on the Internet.  Consumers should explore cloud service provider websites using the existing search engines like (Google and Yahoo) to collect information about all available services and select the best one manually. Unfortunately general purpose search engines are not designed to provide a complete and small set of results that meet the consumer requirements which makes the discovery and selection process a tedious task.

This thesis presents a generic-distrusted framework for cloud services marketplace to: i) automate cloud service discovery and selection process, ii) reduce the time and effort of finding cloud services, iii) make service providers more visible to all consumers, iv) create a shared understanding of cloud service domain and v) improve the overall user experience. In addition, this thesis presents a novel algorithm for cloud services numerical similarity named Percent Distance Similarity (PDS) which is independent of any external values. To overcome the interoperability and vendor lock-in problems, this thesis presents unified cloud services ontology and models the real-life services according to proposed ontology. Finally, this thesis implements two instances of generic framework by adopting two different matching algorithms. First one is based on dominant and recessive attributes algorithm borrowed from gene science and the second one is based on a semantic similarity algorithm and unified cloud service ontology.This thesis is the first attempt to build a cloud service marketplace where cloud service providers and cloud service consumers can trade a cloud service as a utility. It is a global digital market where cloud consumer can find the solution that match his/her needs by simply submitting a request. Comparison done with existing systems on real-life cloud services collected from providers websites based on four parameters (number of matched services, execution time, average Score and recall). Semantic approach based on cloud ontology reduced the execution time by 20% and maintain the same values for all other parameters. On the other hand, Non-semantic approach reduced the execution time by 57% but showed lower value for recall.

S. V Lakshmi , January 2019, PhD Awarded

The thesis work focused on minimizing the tradeoffs that exists
amongst the accuracy, minimum response time and cost of query optimisation techniques using genetic algorithms.
The performance metrics used for the proposed techniques in the thesis are the Average Cost of Query Execution Plan such as Join processing cost, local processing cost and global communication cost based on the accuracies of the executed queries results.