Search

Deep learning has gained much attention and been applied in many different fields. In this paper, we present a web application developed to identify and detect the number of distinct feral cats of Australia using deep learning algorithms targeted to data captured from a set of remote sensing cameras. Feral cat recognition is an especially challenging application of deep learning since the cats are often similar and, in some cases, differ only in very small patterns on their fur. Given the automated, sensor-based image capture from remote cameras, further challenges relate to the limited number of images available. To tackle this, we train four neural network models to distinguish 75 classes (i.e. distinct feral cats) using 30 to 80 images for each class. Based on a range of evaluation metrics, we select Mask R-CNN model with ImageNet pre-trained weights augmented with the ResNet-50 network as the basis for the web application. Using images of cats from cameras in five different forests around Victoria, we achieved an average accuracy of identification of individual (distinct) cats of 89.4% with a maximum accuracy of 96.3%. This work is used to support ecologists in the School of Biosciences at the University of Melbourne.

Understanding the flow of traffic on road networks is increasingly important especially with the continued urbanization of the global population. Numerous hardware and software technologies have been applied to measure traffic volumes by Government agencies and/or organizations such as Google, however they are either expensive to deploy; limited in their ability to disambiguate the kinds of vehicles on the road network, or of increasing importance, they infringe on the privacy of individuals, e.g. tracking phones. In this paper we describe work applying deep learning technologies to identify and classify different vehicles on the road network of Victoria with specific focus on heavy goods vehicles (trucks and trailers). Specifically, we present an approach to automatically detect, classify and count the unique classes of trucks and trailers that are found on the road network and the direction of travel. We apply and compare leading deep learning approaches including You Only Look Once version 3 (YOLOv3) and Single Shot Multi-Box Detector (SSD). This paper builds upon earlier work [1] which focused on data (video) from a single traffic junction in Melbourne. This work is based on a wider range of data (videos) from locations reflecting the diversity of road use including multi-lane motorways, rural roads and city roads.

The Internet has become an essential part of everyday life. It links people with enormous amounts of information covering almost any topic imaginable. However harmful or inappropriate information such as pornography can also be easily found on the web which should not always be available, especially to minors. Internet filters are typically used to block such inappropriate content. These are largely based on the metadata related to the websites or by directly blocking the URLs related to those websites. However seemingly innocuous websites can contain undesirable images that should not be accessible to children. In this paper, we describe how images and videos can automatically be identified (classified) without any human supervision based on their subject matter. To achieve this, we apply deep learning methods to detect and classify adult-only image content from both images and live videos. We use the TensorFlow library and two pre-trained models: MobileNet_v1 and Inception_v3, with an official (academic) pornography dataset including associated labelling. The performance of each model was investigated. The final solution was delivered as an iOS application to detect and classify photos and live videos based on their adult-only content. The app achieved an accuracy of over 92%.

The impacts of the Internet and the ability for information to flow in real-time to all corners of the globe has brought many benefits to society. However, this capability has downsides. Information can be inexact, misleading or indeed downright and deliberately false. Fake news has now entered the common vernacular. In this work, we consider fake news with specific regard to social media. We hypothesise that fake news typically deals with emotive topics that are deliberated targeted to cause a reaction and encourages the spread of information. As such, we explore sentiment analysis of real and fake news as reported in social networks (Twitter). Specifically, we develop an AWS-based Cloud platform utilising news contained in the untrustworthy resource FakeNewsNet and a more trusted resource CredBank. We train algorithms using Naive Bayes, Decision Tree and Bi-LSTM for sentiment classification and feature selection. We show how social media sentiment can be used to improve the accuracy in identification of fake news from real news.

Ethanol is a chemical-colourless compound that widely used in pharmaceutical, medicines, food products, and industrial applications. As the demand for ethanol is rising recently, attention has been given on metabolic engineering of Escherichia coli (E.coli) to enhance its production through alteration of its genetic content. This research mainly aimed to optimize ethanol production in E.coli using a gene knockout strategy. Several gene knockout strategies like OptKnock and OptGene have been proposed previously. However, most of them suffer from premature convergence. Hence, a hybrid of Particle Swarm Optimization (PSO) and Minimization of Metabolic Adjustment (MOMA) algorithm is proposed to identify the list of gene knockouts in maximizing the ethanol production and growth rate of E.coli. Experiment results show that the hybrid method is comparable with two state-of-the-art methods in term of growth rate and production.

© Springer Nature Switzerland AG 2019. Big data analytics typically requires large amounts of resources to process ever-increasing data volumes. This can be time consuming and result in considerable expenses. Analytics-as-a-Service (AaaS) platforms provide a way to tackle expensive resource costs and lengthy data processing times by leveraging automatic resource management with a pay-per-use service delivery model. This paper explores optimization of resource management algorithms for AaaS platforms to automatically and elastically provision cloud resources to execute queries with Service Level Agreement (SLA) guarantees. We present admission control and cloud resource scheduling algorithms that serve multiple objectives including profit maximization for AaaS platform providers and query time minimization for users. Moreover, to enable queries that require timely responses and/or have constrained budgets, we apply data sampling-based admission control and resource scheduling where accuracy can be traded-off for reduced costs and quicker responses when necessary. We conduct extensive experimental evaluations for the algorithm performances compared to state-of-the-art algorithms. Experiment results show that our proposed algorithms perform significantly better in increasing query admission rates, consuming less resources and hence reducing costs, and ultimately provide a more flexible resource management solution for fast, cost-effective, and reliable big data processing.

Social media is one of the primary communication tools for Internet users. Twitter, one of the most popular social medias, has more than one hundred million daily active users. These tweeters tweet a large number of tweets every day containing a rich and diverse collection of information. At the same time, the problem of obesity is becoming a serious issue all over the world. In this paper, we consider the impact of the geographical location of fast food restaurant on the body mass index (BMI) and levels of obesity of individuals in Melbourne through data analytics around social media.

Clustering algorithms are essential for many big data applications involving point-based data, e.g. user generated social media data from platforms such as Twitter. One of the most common approaches for clustering is DBSCAN. However, DBSCAN has numerous limitations. The algorithm itself is based on traversing the whole dataset and identifying the neighbours around each point. This approach is not suitable when data is created and streamed in real-time however. Instead a more dynamic approach is required. This paper presents a new approach, RT-DBSCAN, that supports real-time clustering of data based on continuous cluster checkpointing. This approach overcomes many of the issues of existing clustering algorithms such as DBSCAN. The platform is realised using Apache Spark running over large-scale Cloud resources and container based technologies to support scaling. We benchmark the work using streamed social media content (Twitter) and show the advantages in performance and flexibility of RT-DBSCAN over other clustering approaches.

The effect of seasonal epidemics and potentially pandemics represents a significant issue for public health. In this context, early warnings and real time tracking of the spread of disease is highly desirable. In this paper, we address the problem of detecting disease outbreaks through an automated, scalable Cloud-based system for collecting, tracking and analyzing social media data. Specifically, the focus here is targeted to three prevalent diseases (flu, chickenpox and measles) across three Australian cities using data from the Twitter micro-blogging platform. The epidemics related tweets are extracted using an ensemble learning classifier consisting of a combination of Support Vector Machines, Naive Bayes and Logistic Regression and comparing the results with the Google Trend data to assess the effectiveness of the overall approach.

Semantic annotation framework that allows enriching locations or trajectories with semantic abstractions of the raw spatiotemporal data benefits understanding the semantic behavior of moving objects. Existing semantic annotation approaches mainly analyze specific parts of a trajectory, e.g. stops, in association with data from 3rd party geographic sources, e.g. (POI) points-of-interest, road networks. However, these semantic resources are static thus miss important dynamic event information. Recent location-based social networking provides a new dynamic and prevalent source of human activity data that can be a potential semantic resource for annotation. However, using the large-scale spatiotemporal data from online social media gives rise to privacy concerns. This paper thus presents a privacy-preserving semantic annotation framework P-SAFE that (i) identifies dynamic region of interest (DRI) from large-scale data provided by location based social networks whilst labelling of DRI into appropriate categories derived from spatial and temporal features of geotags, (ii) aligns trajectories to a set of DRI and enriches trajectories with semantics annotation derived from aligned DRI via THMM model, and (iii) embeds robust privacy-preserving mechanisms under differential privacy in each stage that accesses to raw data. P-SAFE approach tackles the privacy and utility trade-offs for meaningful geographic regions identification and labeling as well as trajectory semantic annotation under differential privacy whilst combining them into a single task. We demonstrate the effectiveness of P-SAFE approach on a dataset of large-scale geotagged tweets and a benchmark trajectory dataset for DRI construction and trajectory semantic annotation evaluation. The experimental results illustrate that P-SAFE not only provides robust privacy guarantees but remains approximate 45–56% accuracy for meaningful geographic regions labelling and 62–76% accuracy for trajectory semantic annotation.

Scientific experiments often involve use of shared resources across organization boundaries in distributed collaborative environments. They are more often enabled through web services. A plethora of research is undertaken to protect individual web services. They include centralized security models wherein the main focus is on centralised Virtual Organization (VO) specific attribute authorities, e.g. VOMS, which can be used by collaborative service providers to make authorisation decisions. And a decentralized security model wherein each service provider themselves are responsible for the assignment of roles/privileges to the different members of collaborative environments. Workflows themselves can be orchestrated in centralized or decentralized orchestration models. In this research work we have identified a number of architectural design patterns for security-enabled workflows executions. These patterns are based on the different workflows execution and security models. The key issues in such patterns as well as a rationale of choice are provided. An overview of a security-oriented workflow framework is provided that can tackle some of the issues identified in these patterns.

Scientific workflows are composed of different services to support scientific experiments. Often such services are provided by different organizations that can have their own autonomous access control policies. Workflows are often shared and repurposed with the same and/or different datasets to repeat scientific experiments, therefore, different users can require different privileges to access different services to execute (enact) a given workflow. It can be the case that a given user may not have sufficient privileges to access some of the services of the workflow. As such, it needs to be ascertained whether a user (or enactment engine acting on behalf of a user) with a given set of security credentials should be allowed to enact a workflow and whether this will lead to runtime failure of the workflow. Ideally it should be determined a priori whether a path exists from the root node of the workflow graph to the leaf node, i.e. that it is possible for the workflow to be fully executable or partially executable on the basis of the available credentials of the user. This paper presents an algorithm and its realization that exploits existing workflow patterns to determine the structural path of the workflow whilst checking the availability of credentials at different service points in the workflow path.

Genetic engineering provides methods to modify the genes of microorganisms to achieve desired effects. This can be done for improved organism growth rate or increasing production yield of a desired gene product. Gene knockout is a technique that can improve the specific characteristics of microorganisms by disabling selected sets of genes. However, microorganisms are complex and predicting the effects of gene modification is difficult. Several algorithms have been proposed to support a range of gene knockout strategies, including BAFBA, BHFBA and DBFBA. In this paper, scaling these algorithms and methods to utilise High Performance Computing (HPC) resources have been explored. The applications have been parallelized on HPC and the scalability and performance of these approaches were explored and documented.