Sensor-rich IoT systems are becoming ubiquitous in our lives, from smart wristbands with IMU, to smartphones with depth cameras, to low-cost embedded networked radars.These systems are providing very good alternative ways for human context detection.Yet, making the robust inference from the multi-modality raw sensor data to individual's context in the wild remains difficult.Furthermore, human context may consist of sensitive information, which needs to be protected from malicious attackers.In this talk, I will discuss my group's ongoing research on addressing some of these challenges with example applications in fitness, health and cyber security.
This paper presents a novel method to identify unseen Asian elephants that are not previously captured or identified in available data sets and re-identify previously seen Asian elephants using images of elephant ears, leveraging a semi-supervised learning approach. Ear patterns of unseen elephants are learnt for future re-identification. To aid our process, elephant ear patterns are used as a biomarker to uniquely identify individual Asian elephant, each of which is attached a descriptor. The main challenge is to learn and use a clustering technique to identify new classes (i.e., elephants) in unlabelled elephant ear image sets and leveraging this data in verifying the labelled images. This study proposes a systematic approach to address the problem to uniquely identify elephants, where we developed: (a) a self-supervised learning approach for training the representation of labelled and unlabelled image data to avoid unWanted, bias labelled data, (b) rank statistics for transferring the models’ knowledge of the labelled classes when clustering the unlabelled images, and, (c) improving the identification accuracy of both the classification and clustering algorithms by introducing a optimization problem when training with the data representation on the labelled and unlabelled image data sets. This approach was evaluated on seen (labelled) and unseen (unlabelled) elephants, where we achieved a significant accuracy of 86.89% with an NMI (Normalized Mutual Information) score of 0.9132 on identifying seen elephants. Similarly, an accuracy of 54.29% with an NMI score of 0.6250 was achieved on identifying unseen elephants from the unlabelled Asian elephant ear image data set. Findings of this research provides the ability to accurately identify elephants without having expert knowledge on the field. Our method can be used to uniquely identify elephants from their herds and then use it to track their travel patterns Which is greatly applicable in understanding the social organization of elephant herds, individual behavioural patterns, and estimating demographic parameters as a measure to reducing the human-elephant conflict in Sri Lanka.
Even though there are numerous prior studies concentrating on finding and forecasting vulnerability, the research community have less attention to the aftermath of vulnerability remedies represented in a code. Our primary goal was to understand better what happens to source code following vulnerability remediation via source code and repository analysis. For this research, primarily two distinct data sets were constructed. A dataset containing vulnerability fixes data and a dataset with the data related to the bug fixes. In this context, we have defined the vulnerabilities as the entries that are associated with Common Weakness Enumeration(CWE) entries. Accordingly, in our research, we examined the 80 most popular open-source javascript-based repositories, which contained more than 401K commits, to conduct a timeline analysis based on vulnerability fixes and introductions. Based on these commits, we have extracted source files and generated corresponding Abstract Syntax Tree (AST) for each file. Following that, we have introduced a way to calculate an entropy value on AST and a simple algorithm to identify a significant instance based on that value. Moreover, we have shown that fixing a vulnerability in code makes a considerable change in the code AST than a bug fix by carrying out hypothesis testing. Additionally, we demonstrated that code churn significantly increases following the vulnerability fix.
This study aims to categorize the service types of IP addresses based on running services. The service types can be categorized into Proxy, VPN, Cloud, CDN, and Hosting. The reason for this classification is, that by using these insights one can build better intrusion detection systems. For example, if the IP addresses of the daily users of the websites are mostly VPNs and proxies then owners can be aWare of the traffic and IP addresses of the visitors. Another advantage is one can build better IP blacklists and understand the nature of the ASNs(Some ASNs contain more VPN IPs, and some contain more cloud IPs likewise). The data gathering was done using many methods such as Internet-wide scanning, ASN data collection, and Whois data collection. By using these techniques millions of IP addresses were collected. Many well-known cloud vendors have officially published their IP addresses and this study takes these published IP ranges as the ground truth. About 130,000 IP addresses were collected as the ground truth. By merging these ground truth IPs, with the datasets which are collected by various techniques, the final ground truth IP dataset was constructed. By using this dataset the most important features were identified and trained in various machine learning models. From these trained machine learning models Random Forest Classifier was chosen as the best classmer for this classification since it achieves 94.71% of accuracy and a 95% F1 score.
The human microbiome consists of the microbes that live in and on our body and it plays an important part in our life helping in digestion, developing immunity, etc. There are mounting evidence that shows the relationship between the composition of the gut microbiome and the occurence of various diseases such as type 2 diabetes, colorectal cancer and liver cirrhosis. Machine learning approaches have been used to predict the likelihood of having such diseases using the microbial profile of an individual. In this paper, we explore a deep learning approach to predict health status of an individual utilizing their gut microbial profile. Health state prediction was performed using a dataset containing healthy and non-healthy raw shotgun stool metagenomics samples profiled at species and genera levels. The species level profiles outperformed the genera level profiles in terms of AUC giving a score of 0.957. The best performance for the deep learning approach recorded an AUC score of 0.98 for species level profiles when the sparseness in the datasets were removed. Further, the deep learning approach was able to outperform the state of the art approach.
The rapid progress in deepfake content generation has now come to a point where it raises significant concerns about the implications for society. Therefore, a new challenge of detecting deepfakes arises to protect individuals from potential misuse. Even though introduced detection algorithms perform well on high-quality deepfakes, detecting low-quality deepfakes has been challenging. As a remedy, researchers try to feed more training data to increase detection ability. HoWever, providing more data and processing them is not always feasible in a practical scenario. Thus, for the first time in this domain, we propose to employ super-resolution (SR) as a preprocessing step instead of feeding more data to improve low-quality deepfake detection. Extensive experiments were conducted on the FaceForensics++ deepfake dataset. Initially, three baseline models, Meso-4, MesoInception-4, and XceptionNet, were trained and tested on the dataset without any preprocessing mechanism. XceptionNet outperformed with 90.54% accuracy revealing deeper networks detect low-quality depfakes adequately. Then those baseline models were trained with SR preprocessing. To do that, we employed two SR networks, called VDSR and RSRGAN. RESRGAN+XceptionNet outperformed the previous baseline models by obtaining 96.05% accuracy, showing SR preprocessing usefulness in low-quality deepfake detection. Further experiments utilizing performance metrics, statistical tests, and visualization of activation maps showed that SR preprocessing is promising when applied to deepfake detection networks and detection algorithms experience a significant performance.
Since the beginning of civilization, agriculture has played a significant part in the economy of a nation. Currently, as the population continues to increase at a rapid rate, arable lands are dwindling alongside urbanization. Even though farmers devote a substantial amount of time and effort to farming, environmental factors such as seasonal shifts can have a significant impact on the crop. Smart agriculture is implemented to boost the production of high-quality goods and address the lack of control over the farming process. The intelligent greenhouse technology proposed here is called “GSense,” and it could boost plant productivity by managing the greenhouse’s climate. In addition, this solution is useful for novices who are just beginning out in agriculture because it can make recommendations to its user. The execution of the solution is complemented by a mobile application and a desktop application via which the user may submit inputs and examine real-time sensor data
The Index of Industrial Production (IIP) is a leading indicator used by economic forecasters to predict the overall economic activity of a country. However, no published studies up to date have analyzed and forecasted the IIP of Sri Lanka, which is the only measure of the nation’s physical volume of production. On the other hand, the evolution of information technology has allowed researchers to access thousands of economic series with incredible ease. The usual time series models, such as the widely used Vector Autoregressive (VAR) model can only incorporate a limited number of variables due to the issue of scarce degrees of freedom. Limiting the VAR model to a selected number of variables results in the contamination of economic policy decisions. Dynamic Factor Models (DFMs) are a potential candidate to address this issue by incorporating all available information to the study. However, there is no literature on such applications on a small-scale economy such as Sri Lanka’s to forecast future trends. Therefore, this research aims to follow a dynamic factor approach to forecasting the IIP of Sri Lanka. The objectives of this study are twofold: to explore the different associations between the IIP and other economic indicators, and to investigate how well data-rich DFMs perform compared to classical time series and Long Short-Term Memory networks (LSTMs) when forecasting the IIP. The model estimation was based on the sample from 2000 January to 2018 December and finally, twelve month out-of-sample forecasts were produced for the period until 2019 December. Twenty-three monthly economic variables were considered. The results showed that DFMs, including Factor Augmented VAR (FAVAR) outperformed the traditional VAR, while LSTMs produced the best overall forecasts. Moreover, it was identified that investment imports, mineral and agricultural exports, Colombo Consumer Price Index (CCPI), electricity power consumption and gold price have a significant impact on the IIP.
Sri Lanka is one of the well-known international’s pinnacle tea exporters with a high global demand attracting millions of foreign exchanges, which strengthens the economy of the country. Despite the fact that tea brings a good source of foreign exchange, the tea industry lacks efficiency and effectiveness during the assessment of plucked tea leaves which compromises the significant quality of tea. While studies have revealed various factors affecting the tea quality, key factors are identified as the presence of tea diseases, pest attacks, the mixture of fresh and mature tea leaves, and the mixture of tea grades present in the tea sack. In this paper, we focus on automatically assessing the quality of tea leaves for a single tea leaf and bulk tea leaves before initiating the tea manufacturing process. The proposed tAssessee system allows the user to upload the image of a single tea leaf or bulk tea leaves to automatically assess four different quality factors of tea leaves such as disease, pest attack, freshness, and grade using Convolutional Neural Network based models and using various image processing techniques. This will assist the tea supervisors in the tea factories to automatically assess the quality of tea leaves where the manufacturing process can be segregated according to the quality of tea leaves and determine the pricing accordingly. Extensive experiments performed using the tea leaves images gathered in tea factories reveal, that the proposed tAssessee system can assess the quality of single tea leaf and bulk tea leaves with the accuracy range of 87% - 98% and 91% - 100% respectively.
Communication and presentation of beliefs became easier than in previous decades due to the rapid rise of information technology and computer science. Because social media is accessible worldwide via the internet, anyone can simply target someone or a group who adheres to a different culture or belief. While everyone has the freedom to express their own opinions, it should not be destructive, and everyone has the right to be free of hate speech. Because there are no automatic mechanisms for detecting hate speech on social media, anyone can be readily targeted. Because social media service providers do not have extensive linguistic expertise of some languages, such as Sinhala, it may take a few days for them to delete hate-related comments from the material after they become aware of them. As a result, detecting hate speech in the Sinhala language is an urgent and crucial task. Machine learning and deep learning based algorithms were employed in this study to automatically recognize Sinhala hate speeches broadcast on social media. Bag of words, Tf-idf, Word2Vec, and FastText feature extraction methods were used to extract features from the comments. Logistic Regression, Multinomial Naïve Bayes, Support Vector Machine, XGBoost, Random Forest machine learning models and CNN, RNN, LSTM deep learning models were trained using two pre-collected datasets with different sizes. The best six models were then chosen and test set performances were shown. According to this study, FastText with RNN has the greatest AUC ROC 0.71 with 70% accuracy for the test set.
Road pavement anomalies can result in many negative effects such as damages to vehicles, poor ride quality, additional ride time and road traffic accidents. Therefore it is important to regularly monitor and maintain roads, according to the standards. The traditional methods of road anomaly detection are expensive, time-consuming, and require the supervision of experts. Crowdsourcing systems provide an inexpensive and robust solution to overcome these challenges in traditional approaches. This paper proposes a platform to identify and classify road anomalies from crowdsourced accelerometer smartphone data, adjusting to different vehicle speeds and other characteristics. The collected accelerometer data is preprocessed using noise filtering and reorientation techniques and the anomalies are identified through a fuzzy logic approach and further classified based on the anomaly severity using machine learning models. The results from the conducted experiments suggest that the proposed method is capable of successfully identifying and classifying anomalies from crowdsourced data.
As Majority of the market is presently occupied by Android consumers, Android operating system is a prominent target for intruders. This research shows a dynamic Android malware detection approach that classifies dangerous and trustworthy applications using system call monitoring. While the applications were in the execution phase, dynamic system call analysis was conducted on legitimate and malicious applications. Majority of relevant machine learning-based studies on detecting android malware frequently employ baseline classifier settings and concentrate on selecting either the best attributes or classifier. This study examines the performance of K Nearest Neighbor (KNN), factoring its many hyper-parameters with a focus on various distance metrics and this paper shows performance of KNN before and after performing Principal Component Analysis (PCA). The findings demonstrate that the classification performance may be significantly improved by using the adequate distance metric. KNN algorithm shows decent accuracy and improvement of efficiency such as decreasing the training time After PCA.
The development of Business and Social Computing applications is a challenging task due to the fluidity of requirements. Often requirements evolve throughout the development process. Systems need to meet the needs of multiple user groups, sometimes with conflicting requirements. Thus the best option is to involve the users in the development process at a level they can participate as developers to benefit from their in-depth application domain knowledge. The Meta design paradigm is one of the best approaches for this. With the CBEADS (Component Based E Application Development and Deployment Shell) framework the applications can design in a few clicks, by specifying the application name, and then the functionalities. they have via use-cases, Each use case consists of main functions that which user can perform such as view, create, update, and delete. So the users have the freedom to build things from scratch easily. So the API endpoints will be ready to test the system. This Would help to identify the basic requirements. The newly designed application can be used easily as a prototype system too. Nowadays software development happens as rapid development. So the CBEADS framework is capable of bridging the gap for identification of clear requirements from the users and saving time by increasing the productivity of the developers.
Alzheimer’s disease (AD) is a chronic, incurable disorder that worsens with time and requires early diagnosis in order to treat and manage AD patients. Diffusion MR imaging and structural brain networks provide a great amount of information about the brain that hasn’t been thoroughly investigated before. Graph Neural Networks (GNN) are used to process and learn the graph data structure in deep learning. There hasn’t been any research on using anatomical brain networks With GNN to identify AD so far. In this paper, an efficient GNN architecture is used to classify individuals into Cognitively Normal (CN) and AD subjects using anatomical brain networks as graphs. The input labelled structural brain graphs of CN and AD are used to categorize AD and CN individuals using this GNN architecture. The proposed method is tested using a dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) research, which includes 100 CN and 62 AD subjects. The proposed system effectively demonstrates brain graph properties and provides a reliable Alzheimer’s disease detection classifier. The deep learning system achieves a prediction accuracy of 97 percent, indicating that the proposed classification model is more resilient and perfect than the earlier methods. Our approach changes the way biomarkers of AD are detected and could provide clinicians with more confidence in automated AD diagnostic systems.
Higher education systems worldwide have undergone expansions in line with the increasing efficacy and prominence of cloud computing, leading to enhanced productivity, efficiency, and cost reduction. This study proposes a Comprehensive Technology Readiness Adoption Model in assessing the adoption of cloud computing services and evaluating its contribution to improving literacy levels in higher educational institutes in Nigeria. Primary data was obtained through a survey with a healthy number of samples across several institutions in Nigeria, with the aid of structural equation modeling analysis, the study surveyed 384 staff and students of higher institution in Nigeria. Using statistical techniques including variance-based structural equation modeling, artificial neural network, regression, factor analysis, and correlation, it was found that higher institutions in Nigeria have a positive perception of cloud computing as well as high levels of readiness for adoption. This finding leads to the recommendation that higher institutions in Nigeria should intensify efforts in making wider adoptions of cloud computing for all staff and students in educational settings. The devised model has proved to be effective and accurate as a tool for assessing broad elements concerning the adoption of computing technologies especially under the context of higher education in developing countries.
With the popularity of smartphones, mobile application (A.K.A Mobile App) development has become a booming industry all across the World. One of the main hurdles that app developers are facing, is understanding users’ needs and catering their products to satisfy the users. Though Users are one of the main stakeholders of the App development process it is harder to incorporate them into the requirement elicitation process. Numerous studies have shown that incorporating user reviews in the requirement elicitation process paves the way to a better understanding of user needs which, in turn, helps developers develop better apps that satisfy the targeted audience of the app. In this paper, we introduce a CNN-based approach to analyze user reviews using Aspect-based Sentiment Analysis (ABSA). The results show that our approach could achieve 87.88%, 93.75%, and 31.25% improvements in aspect category classification and 16.43%, 23.35%, and 3.72% improvements in aspect sentiment classification over the baseline results for AWER dataset in productivity, social networking, and game domains respectively.
Due to both false-positive structure identification and flaws in the predicted structures, chemical structure identification from documents remains a complex challenge. Current techniques rely on customized rules and subroutines that, although usually effective, recognition rates are insufficient and systematic improvement is difficult at certain times. Especially when it comes to the identification of hand-drawn Lewis Structures of molecules, most of these existing methodologies fail. Therefore, through this research, we present a system to identify a camera-captured, hand-drawn Lewis structure of a molecule using Machine Learning and Deep Learning concepts. Due to less availability of hand-drawn Lewis structures, we had to make our own dataset and therefore the project was limited to 15 different hydro carbons. Moreover, we provide the users with a mobile application that can visualize the identified molecule in a 3-D space using Augmented Reality. Our machine learning model details are available on the Github (https://github.com/MZJGroup/MoleAR)
Traditional insurance claims processing systems are no match for the modern world due to the increasing population of vehicles and the resulting number of accidents. In this paper, the authors present a novel idea to automate the tedious processes in the insurance industry. The presented system consists of three main components namely, re-identify the make and model of the vehicle, identify the damaged automobile component, type, and severity, and compute an accurate repair estimate using damage component identification. Also, automate the documentation process by identifying the relevant fields in the voice input provided by the user. This ensures both the parties involved in this process will be benefited from the proposed system. Presented solutions Were designed using the aid of Artificial Intelligence techniques, mainly CNN models and Natural language processing techniques.
The Sinhala language is widely used on social media by using the English alphabet to represent native Sinhala words. The standard script of English language is Roman script. Hence we refer to Sinhala texts transliterated using English alphabet as Romanized-Sinhala texts. This process of representing texts of one language using the alphabet of another language is called transliteration. Over the time Sinhala Natural Language Processing (NLP) researchers have developed many systems to process native Sinhala texts. However, it is impossible to use the existing Sinhala text processing tools to process Romanized-Sinhala texts as those systems can only process Sinhala scripts. Therefore these texts need to be transliterated back using their original Sinhala scripts to be processed using existing Sinhala NLP tools. Transliterating texts backwards using their native alphabet is referred to as back-transliteration. In this study, we present a Transliteration Unit (TU) based back-transliteration system for the back-transliteration of Romanized-Sinhala texts. We also introduce a novel method for converting the Romanized-Sinhala scripts into TU sequences. The system was trained using a primary data set and evaluated using an unseen portion of the same data set as well as a secondary data set which represents texts from a different context to the primary data set. The proposed model has achieved 0.81 in BLEU score and 0.78 in METEOR score on the primary data set while achieving 0.57 in BLEU score and 0.47 in METEOR score on the secondary data set.
