Search

The recent surge in the volume of real-time condition-monitoring data generated from connected devices such as from Industrial Internet of Things (IIoT), has enabled wide spread adoption of machine learning techniques to improve the efficiency, effectiveness and intelligence of the industrial process. However, a key obstacle to this whole adoption is that the data is distributed across multiple industrial assets and are usually not shared across the industry partners as they are constrained by complicated administrative process, industry competition concerns and the need to comply to privacy laws. Federated Learning, a privacy preserving machine learning technique coined by Google, is a likely and efficient technique to overcome this obstacle. Federated Learning enables multiple industries to train the condition-monitoring data without revealing their corresponding data and assets. In this chapter, we present our proposed decentralized federated learning based solution and its integration with an IIoT smart manufacturing platform. We demonstrate the effectiveness of our platform with real world datasets and different machine learning algorithms. We observe that the proposed federated learning approach can achieved an equivalent performance compared to the traditional centralized-based learning approach in terms of accuracy and efficiency.

Countries across the globe have been pushing strict regulations on the protection of personal or private data collected. The traditional centralized machine learning method, where data is collected from end-users or IoT devices, so that it can discover insights behind real-world data, may not be feasible for many data-driven industry applications in light of such regulations. A new machine learning method, coined by Google as Federated Learning (FL) enables multiple participants to train a machine learning model collectively without directly exchanging data. However, recent studies have shown that there is still a possibility to exploit the shared models to extract personal or confidential data. In this paper, we propose to adopt Multi Party Computation (MPC) to achieve privacy-preserving model aggregation for FL. The MPC-enabled model aggregation in a peer-to-peer manner incurs high communication overhead with low scalability. To address this problem, the authors proposed to develop a two-phase mechanism by 1) electing a small committee and 2) providing MPC-enabled model aggregation service to a larger number of participants through the committee. The MPC enabled FL framework has been integrated in an IoT platform for smart manufacturing. It enables a set of companies to train high quality models collectively by leveraging their complementary data-sets on their own premises, without compromising privacy, model accuracy vis-a-vis traditional machine learning methods and execution efficiency in terms of communication cost and execution time., This paper appears in the Proceedings of The 20th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing(CCGrid 2020). Please feel free to contact us for questions or remarks

Background: Genotype imputation as a service is developed to enable researchers to estimate genotypes on haplotyped data without performing whole genome sequencing. However, genotype imputation is computation intensive and thus it remains a challenge to satisfy the high performance requirement of genome wide association study (GWAS). Objective: In this paper, we propose a high performance computing solution for genotype imputation on supercomputers to enhance its execution performance. Method: We design and implement a multi-level parallelization that includes job level, process level and thread level parallelization, enabled by job scheduling management, message passing interface (MPI) and OpenMP, respectively. It involves job distribution, chunk partition and execution, parallelized iteration for imputation and data concatenation. Due to the design of multi-level parallelization, we can exploit the multi-machine/multi-core architecture to improve the performance of genotype imputation. Results: Experiment results show that our proposed method can outperform the Hadoop-based implementation of genotype imputation. Moreover, we conduct the experiments on supercomputers to evaluate the performance of the proposed method. The evaluation shows that it can significantly shorten the execution time, thus improving the performance for genotype imputation. Conclusion: The proposed multi-level parallelization, when deployed as an imputation as a service, will facilitate bioinformatics researchers in Singapore to conduct genotype imputation and enhance the association study.

Existing Hadoop MapReduce fault tolerance strategy causes the computing jobs suffering from high performance penalty during failure recovery. In this paper, we propose Fast Recovery MapReduce (FAR-MR) to improve MapReduce performance in failure recovery. FAR-MR includes a novel fault tolerance strategy that combines distributed checkpointing and proactive push mechanism to support fast recovery from task failure and node failure. With distributed checkpointing, computing progress of each task is recorded as checkpoints periodically and kept in distributed data storage. The recovered task can obtain the last progress of the failed task from the distributed storage during failure recovery. In addition, the proactive push mechanism enables the computing results of map tasks to be proactively transmitted to the nodes hosting reduce tasks of the same computing job. When a failure happens, the partial output results being pushed to the reducer nodes can be used by the reduce tasks without the necessity of re-compute. FAR-MR allows a failed task to be recovered efficiently at any node in the cluster. The performance evaluation has shown that the proposed FAR-MR can improve computing job performance by up to 62% and 45% compared to Hadoop MapReduce in the case of task failure recovery and node failure recovery, respectively.

Federated cloud enables data sharing across different organizations. This paper proposes a distributed search index (DS-index) solution to provide distributed search capability in federated cloud, with the target to support multi-attribute range queries with secondary indexes. DS-index deploys a three-layer architecture with a multi-attribute index overlay, a tree-based P2P network layer, and a federated cloud layer. We present a distributed search algorithm with the three-layer architecture. In order to facilitate distributed search, we propose a dynamic mapping algorithm for DS-index to map between different layers. In addition, a Markov Chain-based cost model is defined to facilitate node selection and mapping. With the dynamic mapping algorithm and cost model, the proposed DS-index solution is more cost-effective than the traditional P2P networks as it can reduce the network cost in terms of index maintenance cost and node selection cost. The experiments demonstrate that, our DS-index solution along with its cost model can save the computation resource and reduce network bandwidth consumption by around 30% comparing to the one without cost model. It can also reduce the number of node splits/merges by around 20% comparing to the state-of-the-art solution.

Prompt response to data access requests is the key performance concern of parallel storage systems. Sort partition (SP) is one of the most promising file allocation solutions for the static scenario due to its sorted partition and placing mechanism. The online variant of SP, hybrid partition (HP), however, does not perform well in the dynamic scenario. The recently proposed balanced allocation with sort (BAS) and its online version balanced allocation with sort for batch (BASB) try to compete with SP and HP by providing better load balancing together with sorting. BAS's performance is close to SP, and BASB greatly outperforms HP. In this paper, we proposed new file placement solutions for the parallel storage systems: Optimized Sort Partition (OSP) for the static scenario and Optimized Sort Partition Online (OSPOnline) for the dynamic scenario. By eliminating the drawbacks of SP/HP and BAS/BASB in the file assignment, OSP and OSPOnline can achieve optimized system response time. Conducted simulations show that OSP and OSPOnline steadily outperform their competitors under various conditions.

In this paper, we propose a new reference-based data compression method for efficient compressing of genome sequencing data in FASTQ format. With the advance of the next sequencing technology, the genome data can be generated faster and cheaper, which brings the challenges for efficient storage of these data when used in cloud computing. In order to efficiently store these types of genome data in cloud, content-aware compressing methods have to be developed to make use of the specific file structures. Compared with existing genome-specific compression methods, our proposed content-aware method focused on high compression ratio by taking advantages of repetitive nature of DNA sequence, and using reference genomes in compressing the sequences inside the FASTQ files. The benchmark results of 8 datasets show that our method can achieve highest compression ratio compared with existing FASTQ file compressors.

Federated cloud facilitates data sharing across multiple organizations. Existing distributed storage systems cannot support efficient data search with complicated query requirements. This paper proposes a DS-index solution to provide distributed search in federated cloud, targeting to support multi-attribute range search with secondary index. DS-index includes a three-layer architecture with a multi-attribute index overlay, a tree-based P2P network layer, and a federated cloud layer. In addition, a dynamic mapping algorithm has been presented for DS-index to map between different layers. We have also defined a Markov Chain-based cost model to facilitate node selection and mapping. The proposed DS-index solution is capable to support multi-attribute range queries in the federated cloud. The experiments have shown that, the DS-index solution with cost model can save the computation resource and reduce network bandwidth consumption by around 30% comparing to the one without cost model. It can also reduce the splitting/ merging times by around 20% comparing to the state-of-the-art solution.

The ever increasing amounts of digital data being stored in public and private clouds are challenging users to access and manage the data. With the corresponding storage system reaches Petabyte-scale, or even Exabyte-scale, metadata access will become a severe performance bottleneck. Hence, this paper proposes an efficient multi-dimensional metadata index and search solution for cloud data. By proposing some new mechanism for K-D-B tree based index/search and implementing index partitioning technique, our system can achieve optimized performance in terms of memory utilization and search speed. Experiments show that our system performs much better as compared with existing solutions. In addition, our system can safely scale out in a distributed manner with guaranteed performance.

We experimentally study how to provide differentiated QoP for hybrid packet/WDM networks by intelligently using survivability from packet layer rerouting, WDM layer protection, or cross-layer protection according to different traffic types and survivability requirements.

64 p. Spam messages waste time and resources to the recipients. This dissertation presents the effectiveness of feature selections, particularly,rough set and chi square statistic feature selection methods in combination with J48 decision tree classifier for e-mail classification. Experiments were performed on SpamAssassin corpus, with features selected using word's age, chi square statistic and rough set attribute reduction. Performance is measured based on 10 fold cross validation in terms of Area Under Receiving Operating Characteristic Curve (AUC), precision and recall. The results show feature selection not only can improve the performance of the classifier, but also is a very essential step in e-mail classification. The experiments also reveal that e-mail messages contain a great deal of noise and bad features, which should be removed to increase the performance of the classifier. Master of Science (Smart Product Design)