Abstract

In the realm of biometric security, face recognition technology plays an increasingly pivotal role. However, as its adoption grows, so does the need to safeguard it against adversarial attacks. Attacks involve presenting images of a person printed on a medium or displayed on a screen. Detection of such attacks relies on identifying artifacts introduced in the image during the printing or display and capture process. Adversarial Attacks try to deceive the learning strategy of a recognition system using slight modifications to the captured image. Evaluating the risk level of adversarial images is essential for safely deploying face authentication models in the real world. Among these, physical adversarial attacks present a particularly insidious threat to face antispoofing systems. Popular approaches for physical-world attacks, such as print or replay attacks, suffer from some limitations, like including physical and geometrical artifacts. The presence of a physical process (printing and capture) between the image generation and the PAD module makes traditional adversarial attacks non-viable. Recently adversarial attacks have gained attraction, which try to digitally deceive the learning strategy of a recognition system using slight modifications to the captured image. While most previous research assumes that the adversarial image could be digitally fed into the authentication systems, this is not always the case for systems deployed in the real world. This thesis delves into the intriguing domain of physical adversarial attacks on face antispoofing systems, aiming to expose their vulnerabilities and implications. Our research unveils novel methodologies using white box and black box approaches to craft adversarial inputs capable of deceiving even the most robust face antispoofing systems. Unlike traditional adversarial attacks that manipulate digital inputs, our approach operates in the physical domain, where printed images and replayed videos are utilized to mimic real-world presentation attacks. By dissecting and understanding the vulnerabilities inherent in face antispoofing systems, we can develop more resilient defenses, contributing to the security of biometric authentication in an increasingly interconnected world. This thesis not only highlights the pressing need to address these vulnerabilities but also motivates towards a pioneering approach by exploring simple yet effective attack strategy to advancing the state of the art in face antispoofing security.

Abstract

In the infrastructure industry, the focus is shifting from constructing new suburban buildings to maintaining and rehabilitating existing structures, particularly high-rise buildings. These buildings are prone to various failures due to age, density, and altitude, making regular maintenance vital for safety and longevity. However, traditional inspection methods using heavy machinery and risky rappelling are time-consuming and costly. They often fail to provide comprehensive details for inaccessible surfaces. Finding safer and more efficient inspection approaches is essential to ensure building safety and durability. We introduce an innovative end-to-end pipeline designed specifically for high-rise building inspection. Our proposed method incorporates several key components to ensure effective inspection processes. Firstly, we develop a trajectory generation system for an unmanned aerial vehicle (UAV). This system optimizes the UAV’s trajectory, enabling it to reach the desired destination even in the presence of obstacles. The trajectory can be dynamically adjusted in real-time to ensure efficient navigation. During the UAV’s flight, it captures images of the high-rise building using predetermined camera and drone parameters. These images serve as the basis for building inspections, particularly in detecting cracks. Moreover, the collected pool of images is utilized to construct a detailed 3D mesh model of the high-rise building. This model allows for a comprehensive representation of the structure, facilitating the identification and visualization of detected cracks. By combining trajectory optimization, image capture, crack detection, and 3D mesh modeling, our proposed pipeline offers a comprehensive approach to high-rise building inspection. It presents a promising solution to enhance the efficiency, accuracy, and safety of inspection processes in the field of structural engineering. Our work explores the underexplored domain of deep learning in thermalimaging. It focuses on the classical problem of generating high-resolution images from low-resolution counterparts using Superresolution (SR) techniques. The objective is to extract more details from the original scene by increasing the pixel density, which is particularly valuable in computer vision applications, including pattern recognition and medical imaging. The proposed pipeline aims to achieve real-time video super-resolution using a thermal camera on an embedded edge device.

Abstract

In this thesis, we focus on the secure storage and sharing of Personal Health Records (PHRs) in Internet of Medical Things (IoMT). Due to the high value of personal health information, PHRs are one of the favorite targets of cyber attackers worldwide. Over the years, many solutions, including those based on blockchain, have been proposed; however, most solutions are inefficient for practical applications. For instance, several existing schemes rely on bilinear pairings, which incur high computational costs. To mitigate these issues, we propose a novel PHR-sharing scheme that is dynamic, efficient, and practical. Specifically, we combine searchable symmetric encryption, blockchain technology, and a decentralized storage system known as the Inter-Planetary File System (IPFS) to guarantee the confidentiality of PHRs, verifiability of search results, and forward security. We start by introducing cloud and blockchain and then provide a detailed literature survey of existing works that make use of these technologies for storage and sharing of health records. We then propose our own scheme which provides better security and better efficiency than existing schemes for storage and sharing of PHRs. Moreover, we provide formal security proofs for the proposed scheme and also provide extensive test-bed experimental results which demonstrate that the proposed scheme can be used in practical scenarios related to the IoMT environment.

Abstract

The combination of Internet of Things (IoT) and Machine Learning (ML) technologies has the potential to completely transform air pollution monitoring and management. Low-cost air pollution measurement sensors have grown in popularity in recent years because they enable a low-cost solution to measure air quality in real time. However, these sensors frequently have accuracy concerns and must be calibrated to ensure that their results are valid. This thesis mainly focuses on calibrating low-cost sensors using ML models and the detection of pollutant hot spots in urban areas using IoT devices on a mobile platform. In mobile monitoring of air pollution, IoT-enabled sensors assembled onto automobiles, and portable devices offer real-time data collecting when traveling between areas. This dynamic technique provides useful insights into real-world pollution fluctuations, allowing the identification of pollution sources and trends along traffic routes. The study highlights the importance of calibrating the IoT devices in mobile setting, even when the devices are calibrated in laboratory settings. Real-time ML algorithms can be used to calibrate sensor data based on location, weather, and additional relevant data. In this thesis, a methodology is proposed for detecting emission spikes of PM2.5, CO, and NO2 in polluted urban environments employing portable low-cost sensors. Identification of harmful pollutant concentrations is achieved using two different IoT device types (MegaSense One and Prana Air) mounted on a mobile platform. Reliable identification of the PM2.5, CO, and NO2 emission spikes can be attained by driving through the city on different days. IoT device measurement errors were corrected by ML based calibration against a reference instrument co-located on a mobile platform. ML regression models like simple linear regression (LR), multivariate linear regression (MLR), polynomial regression (PR), support vector regression (SVR), decision tree regression (DT), random forest regression (RF) were applied to calibrate the devices.Among these models RF was the most suitable technique to reduce the variability between the IoT devices due to heterogeneity in the mobile sensing datasets. The spatial variability of PM2.5, CO, and NO2 harmful emission spikes at a resolution of 50 m were identified, but their intensity changes on a daily basis according to meteorological conditions. The data from the PM2.5, CO, and NO2 emission spikes at points of interests that disturb traffic flows clearly show the need for public education about when it is hazardous for persons with respiratory conditions to be outside, as well as when it is unsafe for young children and the elderly to be outside for extended periods of time. This detection strategy is adaptable to any mobile platform used by individuals traveling by foot, bicycle or drones in any metropolis.

Abstract

Biometric technology has long relied on fingerprint recognition as a trusted means of identity verification. While fingerprint enhancement shares similarities with general image denoising tasks, the unique properties of biometric data make it crucial to treat fingerprints differently from typical realworld images. Deep learning methods have the capacity to capture and improve complex patterns and subtle details in fingerprint data, offering a comprehensive solution. This thesis proposes a deep learning-based method, specifically leveraging the U-Net architecture to achieve superior fingerprint enhancement. At its core, the approach employs a multi-task deep network with domain knowledge from minutia and orientation fields for the enhancement of rolled and plain fingerprint images. We investigate different configurations of the U-Net architecture, examine the effects of various architectural elements, and present extensive experimental evidence to support the effectiveness of our proposed approach. Subsequently, we explore a self-supervised learning paradigm with fingerprint biometrics for robust feature representation learning. In this direction, we introduce a pre-training technique by utilizing the learned information from the enhancement task. We adopt the enhancement pre-trained encoder for learning fixed-length fingerprint embeddings. We evaluate the performance of the learned embeddings in the verification task compared to standard self-supervised techniques without the explicit need for enhanced fingerprints. By merging the capabilities of deep learning with the specificity of fingerprint features, the proposed paradigm offers significant improvements in the robustness and accuracy of fingerprint verification. Experimental evaluations on standard benchmarks such as the NIST and FVC datasets demonstrate the efficacy of this approach. We present compelling evidence that our methodology not only improves the quality of fingerprint images but also facilitates a more effective embedding extraction, ultimately leading to enhanced recognition performance. This work highlights the promising role that deep learning, tailored to biometrics, can play in advancing the field of fingerprint recognition. In addition, it opens up avenues for future research in biometrics, particularly in optimizing computational efficiency and in tackling challenges related to partial and distorted fingerprints.

Abstract

Advancements in 3D sensing, driven by the adoption of LiDAR technology, have reignited innovation in autonomous navigation. LiDAR sensors offer real-time, large-scale point clouds, surpassing traditional vision solutions. Datasets like nuScenes and KITTI empower researchers in tasks such as Localization, Place Recognition, and Obstacle Trajectory Prediction. This thesis contributes by exploring the modeling and prediction of large-scale point cloud sequences. Additionally, it showcases a practical application, representing point-cloud sequences as occupancy grid maps and generating trajectories for autonomous navigation. This dual focus enhances our understanding of autonomous navigation complexities, providing valuable insights into real-world implementation challenges. The first study introduces ATPPNet, an innovative architecture tailored to predict future point cloud sequences using Conv-LSTM, channel-wise and spatial attention, and a 3D-CNN branch. Extensive experiments conducted on publicly available datasets demonstrate the model’s impressive performance, outperforming existing methods. The thesis includes a comprehensive ablative study of ATPPNet and an application study showcasing its potential for tasks such as odometry estimation. In the second study, the thesis proposes NeuroSMPC, a novel integration of data-driven frameworks with sampling-based optimal control for real-time applications, particularly onroad autonomous driving. The 3D-CNN layers in NeuroSMPC predicts optimal mean control without iterative resampling, reducing computation time. The approach proves effective in generating diverse control samples around the predicted optimal mean, facilitating real-time trajectory rollout in the presence of dynamic obstacles. The 3D-CNN architecture implicitly learns future trajectories of dynamic agents, ensuring collision-free navigation without explicit future trajectory predictions. Performance gains are demonstrated over multiple baselines in on-road scenes through closed-loop simulations in CARLA. Real-world applicability is showcased by deploying the system on a custom Autonomous Driving Platform (AutoDP). These studies collectively advance the understanding and implementation of autonomous navigation systems in dynamic environments.

Abstract

A Parkinson’s Detection System refers to a set of tools designed to assist in the early diagnosis of Parkinson’s Disease, a progressive neuro-degenerative disease that affects movement control. In literature about Parkinson’s Disease (PD) detection systems, various speech analysis tools were used which included features derived from the voice harmonics, jitter and shimmer, speech rate, etc. While these features describe the changes in the voice patterns of the patient, we found that statistical data of the cepstral coefficients are a more effective classification criteria. This resulted in our use of cepstral coefficients derived from two different methods in our system and the results are compared to a baseline model which uses MFCCs. The first method is focused on the transformation of the speech signal through the time frequency treatment of a wavelet transform to make use of the multi resolution property of the wavelet transform. It is followed by a cepstral analysis in order to extract the cepstral coefficients. The resultant feature dimensions are classified using Support Vector Machine (SVM) with the help of different kernels. This model has been applied to PC-GITA database of diadokinetic word /pa-ta-ka/. The results showed us a performance of 65%. The second method is the Zero-Time Windowing of the speech signal and it allows us a higher spectro-temporal resolution. The cepstral coefficients derived from this signal are used as as the basis for training the model for detection of Parkinson’s Disease. The dataset and classifiers are same as the first method and resulted in 70% when the derivatives of the cepstral features are also used. Both the proposed systems compared favourably to the baseline used which is MFCC based detection system.

Abstract

Creativity is an interesting and exciting topic of research due its mysterious and complex nature. There is an ocean of literature present on the topic and our work is a drop in the ocean. To understand creativity, it is essential to understand the underlying cognitive strategies. Our thesis aims to understand how inducing different types of constraints impacts creative reasoning. To address this question, we conducted a study that examines the role of constraints in welldefined problem-solving in ill-defined problem-solving. We chose variants of Raven’s advanced progressive matrices(APM) for well-defined problem-solving and creative reasoning tasks (CRT) for ill-defined problem-solving. Using traditional APM, we created a novel version of APM with comparatively lesser constraints available to solve the puzzle, called creative APM(cAPM). The cAPM task was designed to induce divergent thinking along with convergent thinking. It is assumed that the difference in constraints changes the nature of the problem space in solving APM and cAPM and may differently affect the following creative reasoning task. We randomly assigned 50 participants to perform APM or cAPM, followed by the CRT, in a fixed order. We observed a significant effect of constraints available to solve well-defined problems on ill-defined problem-solving. The current result showed higher CRT scores when CRT preceded cAPM (Median = 79.25) than APM (Median = 53.00). The result suggests that the flexibility in constraints to solve a well-defined problem induces more divergent thinking alongside convergent thinking and facilitates creative thinking required in ill-defined problem-solving. Following this study, we conducted another study to understand whether the constraints in the well-defined problem solving task impacts the ill-defined problem solving task of a different knowledge domain. The well-defined problem solving tasks remain same as the previous study. For the ill-defined problem solving task we have chosen creative reasoning tasks (CRT) which is of same knowledge domain and alternate uses task[46] of a different knowledge domain. We have collected a sample of 44 participants. We observed a significant effect of the variant of APM on CRT performance. Higher number of rules in the CRT was observed when it was preceded by cAPM (Median = 2) compared to the classic APM (Median = 1). Further, we observed a higher CRT Relationship Score under cAPM (Median = 78.5) compared to the classic APM (Median = 50.5) puzzle conditions. However, the variant of APM did not show a significant effect on AUT, fluency score(p = 0.8 for independent t-test) and elaboration score (p = 0.48 for Mann-Whitney U test). The current results fail to show domain general than specific nature of creative thinking, especially when it is induced varying the constraints of abstract reasoning.

Abstract

Among the many 3D representations, Coordinate-based implicit neural networks or neural fields gained much appreciation in recent times for their ability to represent shape and appearance with very high fidelity and accuracy in 3D computer vision. Despite the advances, however, it remained challenging to build generalizable neural fields for the category of the objects without datasets like shapenet that provide “canonicalized” object instances that are consistently aligned for their 3D position and orientation (pose). Aligning the objects in 3D helps in many tasks for better generalization on 3d scene understanding, classification, and segmentation. 3D pose estimation can also be obtained by aligning the objects in the 3D. There are methods that align 3d objects represented as point clouds/meshes. Now that we have a new promising 3d implicit representation, there is a need to develop a method that helps to align the neural-fields so that we can enjoy the same benefits we had in the space of point clouds/meshes. Unlike point clouds/meshes neural-fields are parametrized by deep neural networks which is very hard to interpret. In this thesis, we present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). Neural-fields, specifically NeRfs describe the 3D scene as a function of density and viewdependent color. Aligning the objects of a category depends on the geometry rather than the color. That’s why CaFi-Net uses density alone to align the objects within the category. Canonicalization is tightly coupled with equivariant networks. In this work, we draw inspiration from 3D Equivariant networks and construct a CaFi-Net as an Equivariant network for rotations. This network directly learns from continuous and noisy density fields by employing a Siamese network architecture. Previous work has done this for points, but handling fields, specifically vector fields, require us to consider rotation equivariance in both the position and orientation of the field. So, to incorporate the rotation equivariance in the fields, we chose the gradient of a scalar field density, which is a vector field, as the signal for building the rotation equivariance in the CaFi-Net. We used spherical harmonics as a basic building block for the equivariant convolution kernels for CaFi-Net. To handle the noisy signal, we weighted the features with the density value at the point. We employed density-based clustering for the segregation of the background and foreground parts, which is utilized in the calculation of the losses. As there is no publicly available dataset, in order to train the CaFi-Net, we created a simulator that renders 54 camera omnidirectional views for 1300 Nerf instances across 13 shapenet object categories. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. As there are no metrics available for canonicalization for neural fields, we used the same metrics used for the point clouds to evaluate the CaFi-Net Performance. Along with the above metrics we have introduced a new metric Ground Truth Equivariance Consistency(GEC) which measures the canonical performance of CaFi-Net to manual labels. Extensive experiments on the above dataset of 1300 NeRF models show that our method matches or exceeds the performance of 3D point cloud-based methods. We conducted ablation studies, which included exploring the choice of the signal, weighing the equivariant features with the density value, assessing the need for the Siamese network, and finally justifying the design choice of the CaFi-Net. In the results section we showed renderings of the Neural-Fields of the object from the canonical pose that are consistent across the category.

Abstract

In the field of robotics, the accurate modeling of uncertainty in the robot’s motion and perception is crucial for effective collision avoidance. Many commodity sensors exhibit nonGaussian noise characteristics, yet existing approaches often assume Gaussian uncertainty to ensure computational tractability. This thesis addresses the gap between non-Gaussian uncertainty and collision avoidance by developing a framework that leverages the distributional characteristics of motion and perception noise. We propose a novel approach that interprets reactive collision avoidance as a distribution matching problem between collision constraint violations and the Dirac Delta distribution. To ensure fast reactivity, we embed each distribution in a Reproducing Kernel Hilbert Space and reformulate the distribution matching as the minimization of the Maximum Mean Discrepancy (MMD). By exploiting the insight that evaluating the MMD reduces to matrix-matrix products, we develop a simple control sampling approach for reactive collision avoidance with dynamic and uncertain obstacles. Furthermore, this thesis advances the state-of-the-art in two key aspects. Firstly, we conduct an extensive empirical study to demonstrate that our planner can effectively infer distributional bias from sample-level information. This insight enables the planner to guide the robot towards good homotopy, utilizing the distributional characteristics of motion and perception noise. In contrast, we highlight how a Gaussian approximation of uncertainty can lead to loss of bias estimation and guide the robot towards unfavorable states with high collision probabilities. Secondly, we compare our proposed distribution matching approach with previous non-parametric and Gaussian approximated methods of reactive collision avoidance. Through tangible comparative advantages, we showcase the superior performance of the distribution matching approach. In summary, this thesis presents a comprehensive framework that addresses the challenge of non-Gaussian uncertainty in collision avoidance. By leveraging the distributional characteristics of motion and perception noise, our approach provides a more accurate and effective method for reactive collision avoidance with dynamic and uncertain obstacles. The empirical study and comparative evaluations demonstrate the advantages of the proposed distribution matching approach over previous methods. This research contributes to advancing the understanding and applicability of collision avoidance strategies, particularly in the context of non-holonomic motion and non-Gaussian uncertainty.