Biblio

With the developments in deep learning, the security of neural networks against vulnerabilities has become one of the most urgent research topics in deep learning. There are many types of security countermeasures. Adversarial examples and their defense methods, in particular, have been well-studied in recent years. An adversarial example is designed to make neural networks misclassify or produce inaccurate output. Audio adversarial examples are a type of adversarial example where the main target of attack is a speech-to-text transcription neural network. In this study, we propose a new defense method against audio adversarial examples for the speech-to-text transcription neural networks. It is difficult to determine whether an input waveform data representing the sound of voice is an audio adversarial example. Therefore, the main framework of the proposed defense method is based on a sandbox approach. To evaluate the proposed defense method, we used actual audio adversarial examples that were created on Deep Speech, which is a speech-to-text transcription neural network. We confirmed that our defense method can identify audio adversarial examples to protect speech-to-text systems.

DPI Management application which resides on the north-bound of SDN architecture is to analyze the application signature data from the network. The data being read and analyzed are of format JSON for effective data representation and flows provisioned from North-bound application is also of JSON format. The data analytic engine analyzes the data stored in the non-relational data base and provides the information about real-time applications used by the network users. Allows the operator to provision flows dynamically with the data from the network to allow/block flows and also to boost the bandwidth. The DPI Management application allows decoupling of application with the controller; thus providing the facility to run it in any hyper-visor within network. Able to publish SNMP trap notifications to the network operators with application threshold and flow provisioning behavior. Data purging from non-relational database at frequent intervals to remove the obsolete analyzed data.

The growing volume of data and its increasing complexity require even more efficient and faster information retrieval techniques. Approximate nearest neighbor search algorithms based on hashing were proposed to query high-dimensional datasets due to its high retrieval speed and low storage cost. Recent studies promote the use of Convolutional Neural Network (CNN) with hashing techniques to improve the search accuracy. However, there are challenges to solve in order to find a practical and efficient solution to index CNN features, such as the need for a heavy training process to achieve accurate query results and the critical dependency on data-parameters. In this work we execute exhaustive experiments in order to compare recent methods that are able to produces a better representation of the data space with a less computational cost for a better accuracy by computing the best data-parameter values for optimal sub-space projection exploring the correlations among CNN feature attributes using fractal theory. We give an overview of these different techniques and present our comparative experiments for data representation and retrieval performance.

We propose that to address the growing problems with complexity and data volumes in HPC security wee need to refactor how we look at data by creating tools that not only select data, but analyze and represent it in a manner well suited for intuitive analysis. We propose a set of rules describing what this means, and provide a number of production quality tools that represent our current best effort in implementing these ideas.