Defense Notices

As semiconductor technologies continue to scale further into the nanometer regime, it is important to study how non-traditional computer architectures may be uniquely suited to take advantage of the novel behavior observed for many emerging technologies. Neuromorphic computing system represents a type of non-traditional architecture encompassing evolutionary. Reservoir computing, a computational paradigm inspired on neural systems, has become increasingly popular for solving a variety of complex recognition and classification problems. The traditional reservoir computing methods employs three different layers – the input layer, the reservoir and the output layer. The input layer feeds the input signals to the reservoir via fixed random weighted connections. These weights will scale the input that is given to the nodes, creating different input scaling for the input nodes. The second layer, which is called the reservoir, usually consists of a large number of randomly connected nonlinear nodes, constituting a recurrent network. Finally, the output weights are extracted from the output layer. Contrary to this traditional approach, the delayed feedback reservoir replaces the entire network of connected non-liner nodes just with a single nonlinear node subjected to delayed feedback. This approach does not only provide a drastic simplification of the experimental implementation of artificial neural networks for computing purposes, it also demonstrates the huge computational processing power hidden in even the simplest delay-dynamical system. Previous implementation of reservoir computing using the echo state network has been proven efficient for channel estimation in wireless Orthogonal Frequency-Division Multiplexing (OFDM) systems. This project aims at verifying the performance of DFR in channel estimation, by calculating its bit error rate (BER) and comparing it with other standard techniques like the LS and MMSE.

Executing high-fidelity tests of radar hardware requires real-time fixed-latency target emulation. Because fundamental radar measurements occur in the time domain, real-time fixed latency target emulation is essential to producing an accurate representation of a radar environment. Radar test equipment is further constrained by the application-specific minimum delay to a target of interest, a parameter that limits the maximum latency through the target emulator algorithm. These time constraints on radar target emulation result in imperfect DSP algorithms that generate spectral artifacts. Knowledge of the behavior and predictability of these spectral artifacts is the key to identifying whether a particular suite of hardware is sufficient to execute tests for a particular radar design. This work presents an analysis of the design considerations required for development of a digital radar target emulator. Further considerations include how the spectral artifacts inherent to the algorithms change with respect to the radar environment and an analysis of how effectively various DSP algorithms can be used to produce an accurate representation of simple target scenarios. This work presents a model representative of natural target motion, a model that is representative of the side effects of digital target emulation, and finally a true HDL simulation of a target.

In the past few years, UAVs have become very popular amongst the average citizen. Much like their military counterpart, these UAVs provide the ability to be controlled by computers, instead of a remote controller. While this may not appear to be a major security issue, the information gained from compromising a UAV can be used for other malicious activities. To understand potential attack surfaces of various UAVs, this paper presents the theory behind multiple possible attacks, as well as implementations of a select number of attacks mentioned. The main objective of this project was to obtain complete control of a UAV while in flight. Only a few of the attacks demonstrated, or mentioned, provide this ability. The remaining attacks mentioned provide information that can be used in conjunction with others in order to provide full control, or complete knowledge, of a system. Once the attacks have been proven possible, measures for proper defense must be taken. For each attack described in this paper, possible countermeasures will be given and explained.

Content Based Image Retrieval also known as QBIC (Query by Image Content) is a retrieval technique where detailed analysis of the features of an image is done for retrieving similar images from the image base. Content refers to any kind of information that can derived from the image itself like textures, color, shape which are primarily global features and local features like Sift, Surf, Hog etc. Content Based image retrieval as opposed to traditional text based image retrieval has been in the limelight for quite a while owing to its contribution in putting away too much responsibility from the end user and trying to bridge the semantic gap between low level features and high level human perception. 
Image Categorization is the process of classifying distinct image categories based on image features extracted from a subset of images or the entire database from each category followed by feeding it to a machine learning classifier which predicts the category labels eventually. Bag of Words Model is a very well known flexible model that represents an image as a histogram of visual patches. The idea originally comes from application of Bag of Words model in document retrieval and texture classification. Clustering is a very important aspect of the BOW model. It helps in grouping identical features from the entire dataset and hence feeding it to the Support Vector Machine Classifier. The SVM classifier takes into account every image that has been represented as a bag of visual features after clustering and then performs quality predictions. In this work we first apply the Bag of Words on well known datasets and then obtain accuracy parameters like Confusion Matrix, MCC, (Matthews Correlation Coefficient) and other statistical measures. For Feature selection we considered SURF Features owing to their rotation and scale invariant characteristics. The model has been trained and applied on two well known datasets Caltech 101 and Flickr- 25K followed by detailed performance analysis in different scenarios. 

Biometric ear authentication has received enormous popularity in recent years due to its uniqueness for each and every individual, even for identical twins. In this paper, two scale and rotation invariant feature detectors, SIFT and SURF, are adopted for recognition and authentication of ear images. An extensive analysis has been made on how these two descriptors work under certain real-life conditions; and a performance measure has been given. The proposed technique is evaluated and compared with other approaches on two data sets. Extensive experimental study demonstrates the effectiveness of the proposed strategy. Robust Estimation algorithm has been implemented to remove several false matches and improved results have been provided. Deep Learning has become a new way to detect features in objects and is also used extensively for recognition purposes. Sophisticated deep learning techniques like Convolutional Neural Networks(CNNs) have also been implemented and analysis has been done.Deep Learning Models need a lot of data to give a good result, unfortunately ear datasets available publicly are not very large and thus CNN simulations are being carried out on other state of the art datasets related to this research for evaluation of the model.

Coherent Raman scattering (CRS) is an appealing technique for spectroscopy and microscopy, due to its selectivity and sensitivity. We designed and built single-fiber-laser-based coherent Raman scattering spectroscopy and microscopy system which can automatically maintain frequency synchronization between pump and Stokes beam. The Stokes frequency shift is generated by soliton self-frequency shift (SSFS) through a photonic crystal fiber. The impact of pulse chirping on the signal power reduction of coherent anti-Stokes Raman scattering (CARS) and stimulated Raman scattering (SRS) have been investigate through theoretical analysis and experiment. 

Our multimodal system provides measurement diversity among CARS, SRS and photothermal, which can be used for comparison and offering complementary information. Distribution of hemoglobin in human red blood cells and lipids in sliced mouse brain sample have been imaged. Frequency and power dependency of photothermal signal is characterized. 
Based on the polarization dependency of the third-order susceptibility of the material, the polarization switched SRS method is able to eliminate the nonresonant photothermal signal from the resonant SRS signal. Red blood cells and sliced mouse brain samples were imaged to demonstrate the capability of the proposed technique. The result shows that polarization switched SRS removes most of the photothermal signal. 

The main focus of the project is to discuss mining of incomplete data which we find frequently in real-life records. For this, I considered the probabilistic approximations as they have a direct application to mining incomplete data. I have examined the results obtained from the experiments conducted on eight real-life data sets taken from University of California at Irvine Machine Learning Repository. I also investigated the properties of singleton, subset, and concept approximations and corresponding consistencies. The main objective was to compare the global and local approximations and generalize the consistency definition for incomplete data with two interpretations of missing attribute values: lost values and "do not care" conditions. In addition to this comparison, the most useful approach among singleton, subset and concept approximations is also tested for which the conclusion is the best approach would be selected with the help of tenfold cross validation after applying all three approaches. Also it’s shown that even if there exist six types of consistencies, there are only four distinct consistencies of incomplete data as two pairs of such consistencies are equivalent.

With the advent of Internet, the data being added online have been increasing at an enormous rate. Though search engines use information retrieval (IR) techniques to facilitate the search requests from users, the results may not always be effective or the efficiency of results according to a search query may not be high. The user has to go through certain web pages before getting at the web page he/she needs. This problem of information overload can be solved using automatic text summarization. Summarization is a process of obtaining an abridged version of documents so that user can have a quick understanding of the document. A new technique to produce a summary of an original text is investigated in this project. 
Email threads from the World Wide Web consortium’s sites (W3C) corpus are used in this system.Our system is based on identification and extraction of important sentences from the input document. Apart from common IR features like term frequency and inverse document frequency, novel features such as Term Frequency-Inverse Document Frequency,subject words, sentence position and thematic words have also been implemented. The model consists of four stages. The pre-processing stage converts the unstructured (all those things that can't be so readily classified) text into structured (any data that resides in a fixed field within a record or file). In the first stage each sentence is partitioned into the list of tokens and stop words are removed. The second stage is to extract the important key phrases in the text by implementing a new algorithm through ranking the candidate words. The system uses the extracted keywords/key phrases to select the important sentence. Each sentence is ranked depending on many features such as the existence of the keywords/key phrase in it, the relation between the sentence and the title by using a similarity measurement and other many features. The third stage of the proposed system is to extract the sentences with the highest rank. The fourth stage is the filtering stage where sentences from email threads are ranked as per features and summaries are generated. This system can be considered as a framework for unsupervised learning in the field of text summarization. 

“Flight-search” application is an Angular JS application implemented in a client side architecture. The application displays the flight results from different airline companies based on the input parameters. The application also has custom filtering conditions and custom pagination, which a user can interact with to filter the result and also limit the results displayed in the browser. The application uses QPX Express API to pull data for the flight searches.

In Data Mining, classification of data is a challenging task. One of the most popular techniques for classifying data is decision tree induction. In this project, two decision tree generating algorithms CART and C4.5, using their original implementations, are compared on different numerical data sets, taken from University of California Irvine (UCI). The comparative analysis of these two implementations is carried out in terms of accuracy and decision tree complexity. Results from experiments show that there is statistically insignificant difference(5% level of significance, two-tailed test)between C4.5 and CART in terms of accuracy. On the other hand, decision trees generated by C4.5 and CART have significant statistical difference in terms of their complexity.