Electrical and Electronic Engineering - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemNovel Defenses Against Data Poisoning in Adversarial Machine Learning( 2019)Machine learning models are increasingly being used for automated decision making in a wide range of domains such as security, finance, and communications. Machine learning algorithms are built upon the assumption that the training data and test data have the same underlying distribution. This assumption fails when (i) data naturally evolves, causing the test data distribution to diverge from the training data distribution, and (ii) malicious adversaries distort the training data (i.e., poisoning attacks), which is the focus of this thesis. Even though machine learning algorithms are used widely, there is a growing body of literature suggesting that their prediction performance degrades significantly in the presence of maliciously poisoned training data. The performance degradation can mainly be attributed to the fact that most machine learning algorithms are designed to withstand stochastic noise in data, but not malicious distortions. Through malicious distortions, adversaries aim to force the learner to learn a model that differs from the model it would have learned had the training data been pristine. With the models being compromised, any systems that rely on the models for automated decision making would be compromised as well. This thesis presents novel defences for machine learning algorithms to avert the effects of poisoning attacks. We investigate the impact of sophisticated poisoning attacks on machine learning algorithms such as Support Vector Machines (SVMs), one-class Support Vector Machines (OCSVMs) and regression models, and introduce new defences that can be incorporated into these models to achieve more secure decision making. Specifically, two novel approaches are presented to address the problem of learning under adversarial conditions as follows. The first approach is based on data projections, which compress the data, and we examine the effect of the projections on adversarial perturbations. By projecting the training data to lower-dimensional spaces in selective directions, we aim to minimize the impact of adversarial feature perturbations on the training model. The second approach uses Local Intrinsic Dimensionality (LID), a metric that characterizes the dimension of the local subspace in which data samples lie, to distinguish data samples that may have been perturbed (feature perturbation or label flips). This knowledge is then incorporated into existing learning algorithms in the form of sample weights to reduce the impact of poisoned samples. In summary, this thesis makes a major contribution to research on adversarial machine learning by (i) investigating the effects of sophisticated attacks on existing machine learning models and (ii) developing novel defences that increase the attack resistance of existing models. All presented work is supported by theoretical analysis, empirical results, and is based on publications.
-
ItemBig data clustering for smart city applications( 2016)The Internet of Things (IoT) infrastructure for the creation of smart cities consists of internet connected sensors, devices and citizens. This IoT infrastructure generates an enormous amount of data in the form of city-scale physical measurements and public opinions, constituting big data. Smart cities aim to efficiently use this wealth of data to manage and solve the problems faced by modern cities for better decision making. However, interpretation of the massive amount of smart city generated big data to create actionable knowledge is a challenging task. Aggregation and Summarization (data clustering) is a useful tool to create knowledge from raw data from different sources. However, traditional data clustering algorithms are not suitable for unlabelled smart city data owing to its high volume and generation velocity and limited experience about generating phenomenon. This thesis presents a novel framework for clustering tendency assessment for big data: clusiVAT, which provides an aggregated view of the big data to create actionable knowledge. clusiVAT intelligently selects a small number of samples from the data such that the samples retain the approximate geometry of the big dataset. The reordered dissimilarity image of the samples generated using single linkage minimum spanning tree (MST) suggests the number of clusters in the data, which is required as an input for most popular clustering algorithms. The cluster labels are then extended to the non-sampled points using the nearest prototype rule. The clusiVAT framework was applied to two real life smart city applications to understand the underlying patterns hidden in the huge volumes of data to generate knowledge. The first application used clusiVAT for clustering and anomaly detection from the pedestrian and vehicle trajectories obtained from a video surveillance system. Experiments were performed on a real-life MIT trajectories dataset of vehicles and pedestrians from a parking lot scene. The trajectory clusters and anomalies thus obtained were helpful in the high-level interpretation of a scene (crowd behavior modeling), as feedback for a low-level (individual) tracking and activity prediction system and as an alarm for human supervisor. For the second application, clusiVAT was used to cluster large scale (of the order of millions) vehicular trajectories obtained from the GPS traces of taxis in the city of Beijing and Singapore using a novel Dijkstra-based dynamic time warping distance measure. The results facilitated the understanding of spatial and temporal patterns in trajectories and were of great significance for decision-makers to understand road traffic conditions and to propose metro bus corridors and light rail systems for better public transport. Another prominent data generated by smart city IoT infrastructure are high-velocity data streams. Automatic interpretation of these evolving big data is required for timely detection of unusual events. This thesis presents a computationally efficient 'hot' update approach for incremental visualization of evolving cluster structure in streaming data. The new algorithms were demonstrated for two applications: online anomaly detection and sliding window based clustering of time series data. Numerical experiments on weather monitoring data from great barrier reef and the city of Melbourne provided visual clues to the onset of the new structure in streaming data.