Abstract

Contactless fingerprint recognition provides a number of benefits in terms of hygiene and minimizing
sensor-induced artifacts, over traditional contact-based systems, but it is also associated with a number
of challenges that are currently hindering its adoption. The 2D representation of 3D features, in combination with unconstrained illumination, affects the quality of ridges, resulting in inconsistent features.
Also, the lack of interaction information in contactless recognition makes it more prone to presentation
attacks. Current techniques for acquisition, enhancement, and detection of spoofing attacks are mostly
independent and focused on single images, without exploiting any complementary information for improved recognition. As the trend towards contactless biometric recognition is becoming increasingly
necessary for usability and cost-effectiveness in terms of deployment in mobile and security-critical
applications, lower matching accuracy and increased susceptibility to spoofing attacks are significant
challenges. This highlights the need for a method that jointly addresses enhancement, recognition, and
spoof detection while improving cross-modal interoperability.
In our work, we present a unified framework for contactless fingerprint recognition by exploiting
the flash-non-flash acquisition pair as a lightweight active sensing mechanism that extracts complementary illumination information. We thus present the Flash-Non-Flash Fingerphoto (FNF) Database for
a thorough analysis of the illumination impact, and Fusion2Print (F2P), an end-to-end deep learning
framework that learns the optimal enhancement in the spatial, frequency, and color domains to enhance
the quality of the ridge-valley representations and cross-domain recognition accuracy. Furthermore, we
investigate a method for the detection of presentation attacks by exploiting the illumination-dependent
photometric differences between genuine and spoofed fingerprints, and a low-cost 3D pipeline for the
reconstruction of the fingerprint surface geometry from just two images, facilitating the simulation of
digital spoof attacks.
Experimental results show that all components benefit from improvements. F2P obtains AUC of
0.999 and EER of 1.12%. The illumination-aware technique separates real and spoofing samples effectively, and the 3D reconstruction pipeline rebuilds finger geometry and ridge-valley information with
improved contrast. These components constitute a refined unified framework for contactless fingerprint
recognition.

Abstract

Two-wheelers are among the most widely used modes of transportation in many regions of
the world, particularly in developing countries. However, the lack of structural protection for
riders makes them highly vulnerable to road accidents. Understanding and recognising driving
events from vehicle motion data is therefore an important step toward developing intelligent
safety systems, rider assistance technologies, and accident detection mechanisms.
This thesis investigates time-series based deep learning approaches for two-wheeler driving
event recognition using inertial sensor data. A multivariate time-series dataset was collected
using an embedded sensing platform consisting of a tri-axial accelerometer, gyroscope, and GPS
module mounted on a motorcycle. Unlike many existing datasets, the collected data incorpo-
rates diverse riding conditions, including recordings from multiple riders, variations in riding
behaviour, and both daytime and nighttime sessions, providing a more realistic representation
of real-world riding dynamics.
Several temporal deep learning architectures are evaluated for driving event classification, in-
cluding Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), attention-enhanced
recurrent networks, Temporal Convolutional Networks (TCN), and transformer-based models.
To capture both short-term motion patterns and long-range temporal dependencies, a Time-
Series Transformer (TST) architecture is explored and compared against these baseline models.
All models are trained and evaluated using a consistent preprocessing pipeline and hyperpa-
rameter optimisation framework.
Experimental results show that transformer-based temporal modelling achieves competitive
or superior performance compared to recurrent and convolutional approaches, particularly for
challenging manoeuvre-related events such as turns, stops, and bumps. To evaluate practical
feasibility, the trained models are deployed on a Raspberry Pi edge platform, where inference
latency, memory usage, and real-time processing capability are analysed.
The results demonstrate that accurate driving event recognition can be achieved using
lightweight temporal models operating directly on raw sensor streams, enabling the development
of practical on-device safety monitoring systems for two-wheelers.

Abstract

The rapid evolution of modern professional and educational environments has precipitated a global
transition toward a predominantly sedentary lifestyle. In industrialized and developing nations alike,
individuals now spend upward of ten hours daily in a seated position; Toko et al. called it out as "the
modern epidemic". While the digitalization of labor has enhanced productivity, it has simultaneously
introduced significant risks to musculoskeletal and metabolic health. Modern workstations, despite their
ergonomic advancements, facilitate prolonged static loading that the human musculoskeletal system is not
biologically designed to sustain. As desk jobs are becoming more normalized, making people comfortably
numb. Thus, the development of intelligent, non-intrusive, and reliable systems for monitoring and
classifying sitting posture has emerged as a critical frontier in proactive healthcare, workplace ergonomics,
and personal well-being. This is referred to as Sitting Posture Recognition (SPR). Researchers have tried
many different approaches and deployed a diverse range of technologies to build systems for Human
Activity Recognition (HAR) and SPR. These include but are not limited to RGB imaging and feature
extraction, depth sensing and 3D joint tracking, wearable inertial sensors, smart textiles, and pressuresensitive mats (PSMs). PSMs have become increasingly popular as they address several drawbacks
and concerns with the other technologies-they are non-intrusive, generalizable to multiple users, do not
invade privacy by involving a camera, aren't limited by objects blocking a camera FOV, do not meddle
with general bodily movement, etc.
Many different studies have tried different Machine Learning (ML) and Deep Learning (DL) techniques for SPR, and in order to do that, researchers have created their own datasets. However, one of the
primary concerns with existing research has been lack of generalizability due to not having a diverse
dataset that can be a good proxy for the general population. This thesis outlines the creation of a diverse
dataset from 33 different participants, both male and female, with a wide range of physiological attributes
such as height, weight, and age. The data comprises pressure readings from a high-resolution PSM. With
over 40,000 datapoints, a thorough data analysis is performed, and then several ML and DL techniques
are applied to evaluate the performance of different models for SPR. It is found that traditional models
fail to surpass a particular accuracy level, and thus, a Multi-Scale Spatio-Temporal Ensemble Method
with Physics-Informed Features is proposed, which outperforms conventional models

Abstract

Many real-world relational datasets are naturally represented as graphs, but the relationships they
contain are often not observed with complete certainty. In biological interaction networks, for example,
an edge may be supported by experimental evidence, database integration, or computational prediction, and is therefore more appropriately associated with a confidence value than treated as certainly
present or absent. This motivates the use of uncertain graph transactional databases, where each transaction is an uncertain graph and each edge is associated with an existence probability. Mining recurring
subgraphs from such databases is useful for identifying structural patterns that persist across uncertain
graph collections, but it also changes the meaning and cost of support computation.
In certain graph transactional databases, frequent and top-k subgraph mining have been studied extensively, and the support of a pattern is computed by counting the number of graphs in which it occurs.
In uncertain graph transactional databases, however, subgraph occurrence is probabilistic because each
uncertain graph may realise into multiple certain graphs depending on which uncertain edges exist.
Support must therefore be evaluated through expected support rather than deterministic frequency, and
computing the occurrence probability of a subgraph becomes substantially more expensive. Existing deterministic top-k subgraph mining methods avoid the difficulty of selecting a minimum-support threshold, but they assume certain graph transactions. Existing uncertain frequent subgraph mining methods
account for edge uncertainty, but they are primarily formulated as threshold-based mining problems.
The problem addressed in this thesis lies at the intersection of these two settings: mining top-k subgraphs from uncertain graph transactional databases using expected support as the ranking criterion.
This thesis formulates the top-k uncertain subgraph mining problem without relying on a userspecified minimum-support threshold and develops two algorithmic frameworks for solving it. The first
is an exact method, called Top-k Uncertain Subgraph Miner (TUSM), which explores the candidate subgraph space using best-first search and computes exact expected support through occurrence-probability
evaluation. Since occurrence-probability computation is related to DNF counting, the exact method can
become computationally expensive when candidate patterns have many embeddings. To reduce this
cost, the thesis also proposes an approximate method, called Approximate Top-k Uncertain Subgraph
Miner (ATUSM), which replaces exact support evaluation with a sampling-based approximation procedure. Because the approximate setting yields support intervals rather than exact values, ATUSM uses
interval-based comparison and probabilistic distribution queues to maintain candidate rankings during
search

Abstract

This thesis develops and analyzes an adaptive framework for coordinated dual-arm manipulation that
integrates reinforcement learning with optimization-based variable impedance control. Building on the
DA-VIL approach, the work addresses the core challenge of dual-arm tasks: maintaining stable, precise
coordination under changing object properties and strongly coupled arm-object dynamics.
The proposed pipeline uses a learned policy (trained with PPO) to predict time-varying Cartesian
stiffness parameters, and an optimization (QP)-based controller that converts those impedance settings
into joint accelerations and torques while enforcing physical constraints (joint limits, velocity/torque
bounds, and end-effector geometry).
The thesis presents design choices for the policy and reward structure (including an EMA regularizer
for smooth stiffness trajectories), the QP formulation for simultaneous impedance and postural tasks,
and extensive MuJoCo evaluations on a diverse set of large objects with varying masses and geometries.
Quantitative results show consistent reductions in object trajectory-tracking error and improved robustness compared to baseline controllers (fixed impedance, optimization-only, and straightforward
RL+IC). Ablations highlight the importance of smooth stiffness predictions and the combined RL+QP
architecture. The thesis concludes with a discussion of limitations, potential extensions to assembly and
obstacle-aware planning, and plans for sim-to-real transfer and hardware validation.

Abstract

This study seeks to examine India's position within the global data economy, using three analytical
approaches. First, it assesses the extent to which dependency theory and world systems theory, originally
developed to analyse colonial and post-colonial economic relationships, may provide a useful lens for
understanding relationships between actors in the contemporary global digital economy, with data as a
central resource. Within this framework, it characterises the 'core', 'periphery', and 'semi-periphery' of
the data economy, and identifies attributes that would place India within the semi-periphery. This 'middle' position is examined further by exploring India's regulatory approach to cross-border data flows, as
reflected in its domestic legislation and international trade negotiations. The study characterises this as
a 'quasi-localisation' approach, placing onus on the Government of India to flexibly decide the kinds of
data, recipient countries, and transferring entities involved in permitted data transfers. Hence, in such
a context, it would be useful to analyse the government's priorities over time with respect to India's
domestic and international positions in the digital landscape. Thus, the study finally traces how the
Government of India's discourse around the digital economy has evolved over the period 2015 to 2025
within official government communications, and what this may indicate about the priorities and orientation of domestic digital policy.
The study finds that the global data economy appears to exhibit characteristics broadly consistent
with a core-semiperiphery-periphery framework, in which a small number of actors capture a disproportionate share of economic value from global data flows. Countries that exemplify the core are capable
of exerting a significant degree of structural control over other actors through technical, regulatory, and
policy standards, and are able to retain most of the value generated in the landscape. Countries positioned as semi-peripheral actors exhibit growing domestic technological capabilities and a degree of
strategic positioning in trade and regulatory negotiations, while remaining dependent on core actors for
key aspects of digital infrastructure. Some countries that may be analysed to be peripheral actors, are
dependent on core countries for fundamental infrastructure through various mechanisms. Through this
study, India is evaluated to occupy a place in the semi-periphery of the global data economy.
India's domestic data governance approach is characterised in this study as one of 'quasi-localisation',
which is consistent with its 'middle' semi-peripheral position globally. This characterisation was made
observing India's position as neither one of strict data localisation nor of complete support for the free
flow of data, but rather a strategically ambiguous intermediate stance that preserves substantial discre tionary power for the Government of India. This characterisation draws on an analysis of successive
versions of the Digital Personal Data Protection (DPDP) draft bills, the DPDP Act 2023, and the DPDP
Rules 2025, as well as India's strategic participation in international trade negotiations and agreements
with digital trade provisions.
The question of discretionary power to the Government of India lends itself to a consideration of
the evolving priorities of the government over time, for which this study undertakes a computational
text analysis of 8,532 Press Information Bureau press releases, selected from a total corpus of 1,33,902
collected over the period 2015 to 2025. The time period starts from the year of the launch of the Digital
India Mission, and concludes in the year of the release of the DPDP Rules, and hence provides a broad
overview of various developments in India's digital landscape. The analysis finds that certain subthemes
consistently dominated government communications throughout the period, namely those concerned
with national tech missions, while privacy and data flows related discourse has remained considerably
smaller in volume, albeit growing from 2022 onward. The largest volume of discourse happens at the
scale of the 'nation', with the 'local' and 'international' scales seeing a smaller amount of activity. A
small number of ministries, principally the Prime Minister's Office and the Ministry of Electronics and
Information Technology, dominate across subthemes and scale frames, however the presence of other
ministries points to a growing impact of the digital landscape on India's overall governance and its discourse.
Overall, the study locates India as a semi-peripheral actor in the global data economy, characterising its domestic data governance approach as a 'quasi-localisation' approach, and identifies themes of
growth and development as dominant within the domestic governance discourse

Abstract

Large Language Models (LLMs) produce fluent, coherent text across a wide range of Natural Language Generation (NLG) tasks, but this fluency often masks a critical failure mode: hallucination,
where the model generates content that is factually incorrect, semantically unfaithful, or insufficiently
grounded in the available context. Despite growing interest in hallucination detection, two key limitations persist in the literature. First, most detection methods rely on large generative models as evaluators,
introducing high computational cost, limited interpretability. Second, and more fundamentally, existing benchmarks are overwhelmingly English-centric and general-domain, leaving open the question of
whether current detectors generalize to multilingual and domain-specific settings.
This thesis addresses both limitations. We begin by observing that hallucination, when viewed
through the lens of task semantics, naturally reduces to a problem of inferential grounding: a faithful output is one that stands in the appropriate entailment relationship with its source. Building on this
observation, we propose a lightweight, zero-shot hallucination detection framework based on pretrained
Natural Language Inference (NLI) models, which achieves competitive performance on the SHROOM
benchmark without any task-specific fine-tuning or expensive LLM-based evaluation. This serves as
both a practical detection tool and a conceptual foundation.
However, the main problem this thesis establishes to tackle is to propose a domain-specific, multilingual benchmark for the current hallucination detection models, hence, we introduce CAP (Confabulations from ACL Publications), a multilingual benchmark for scientific hallucination detection. CAP
is grounded in 293 award-winning ACL Anthology papers and the dataset spans evaluation across nine
languages -- English, French, Hindi, Italian, Spanish, Bengali, Gujarati, Malayalam, and Telugu --
covering a wide range of resource conditions. A key design decision is the separation of factual correctness from fluency in annotation, enabling controlled evaluation of hallucination independent of surface
quality.
Benchmarking six state-of-the-art detectors on CAP reveals that no existing system exceeds a MacroF1 of 0.46, and performance degrades substantially in low-resource languages and scientific domains.
Further analysis shows that question type is a statistically significant predictor of hallucination rate, a
finding with direct practical implications for how questions should be framed when eliciting factual
information from LLMs.
Together, these contributions argue that progress in hallucination detection requires both task-aware
formulations and linguistically diverse, domain-specific evaluation resources.

Abstract

The rapid adoption of large language models (LLMs) in applications such as recruitment, content
generation, and decision support has raised critical concerns regarding fairness and bias in automated
text generation. This thesis presents a comprehensive investigation into gender and occupational bias
in LLM-generated narratives, combining theoretical formulation, large-scale empirical analysis, and
multilingual evaluation across three complementary studies.
At the foundational level, this work formalizes bias in generative AI using a statistical framework,
defining bias as the deviation between generated outputs and an ideal, unbiased representation. Building
on this, a robust bias quantification pipeline is introduced, incorporating measures such as mean bias
(MB), mean absolute bias (MAB), sentiment analysis, and probabilistic distribution distances. Using
a large dataset of over 11,000 AI-generated job narratives, the study demonstrates that LLM outputs
exhibit measurable and consistent gender bias, particularly in the association of male entities with highintensity, high-stress, and challenging occupations.
Extending beyond single-model analysis, the thesis further explores bias across multiple LLMs,
including Gemini, Mistral, LLaMA, and OpenAI models, and introduces a structured evaluation framework based on narrative sentiment, syntactic agency, and a derived prestige score. The findings reveal
that gender bias persists across all models, though its magnitude and direction vary depending on model
architecture and prompt design. Notably, prompt framing is shown to significantly influence bias, with
negative or incompetence-related prompts consistently amplifying male representation.
A key contribution of this work is its multilingual analysis of bias in English and Telugu, addressing
a gap in existing research that predominantly focuses on high-resource languages. The results highlight
that while bias is present across both languages, its manifestation differs structurally: English narratives
show greater sentiment variability, whereas Telugu narratives exhibit higher agency representations.
These findings underscore the importance of linguistic context in shaping model behavior.
Overall, this thesis contributes a unified, scalable, and interpretable framework for analyzing bias
in generative AI systems. By integrating statistical rigor, large-scale datasets, cross-model comparison,
and multilingual evaluation, the work advances understanding of how bias emerges and propagates in
LLM-generated text and provides a foundation for developing more fair, transparent, and inclusive AI
systems.

Abstract

Climate change discourse varies across contexts, influenced by political priorities, developmental
goals, and social conditions. One important aspect of this variation is the divide between the Global
North and Global South, which shapes how conversations about responsibility, equity, and climate action
are understood and debated. However, this dimension remains insufficiently explored in computational
analyses of online communication. To address this gap, this thesis examines how these variations appear
in platform-mediated discourse on YouTube and extends the analysis to institutional climate narratives
within the Indian think tank space.
Using a mixed-methods computational approach, the study analyzes a dataset of YouTube video
transcripts and their associated comments, collected around key moments in global climate governance,
including major international agreements and conferences. Topic modeling, sentiment analysis, and
large language model (LLM)-based stance detection are used to examine how climate issues are framed
in video content and how audiences engage with these narratives. The findings reveal systematic differences between Global North and Global South discourse. Content from the Global North tends to
emphasize technological solutions, emissions reduction, and policy frameworks, while Global South
narratives highlight development priorities, vulnerability, and questions of equity and historical responsibility. Audience engagement further reflects these divides, with high levels of polarization shaping
how narratives are received and debated.
To complement this analysis, the study examines institutional climate discourse through Indian think
tank publications, focusing on how the concept of a "just transition" is framed within policy contexts.
Rather than directly engaging with the Global North-South divide, these documents frame climate concerns through domestic priorities such as development, energy transitions, and internal inequalities. The
analysis examines how different dimensions of climate justice, including distributive, reparative, representative, and restorative justice, are represented, offering insight into how justice is articulated within
national policy discourse.
By bringing together digital and institutional analyses, this study demonstrates that climate discourse
is not uniform but is shaped by platform structures, geopolitical contexts, and policy environments.
While the Global North-South divide is more visible in online narratives, institutional discourse focuses
on domestic concerns. Overall, the findings highlight the importance of context-sensitive approaches
to studying climate communication and offer insights into how narratives of responsibility, equity, and
climate action are constructed and contested across different domains.

Abstract

Robotic manipulation of large and complex objects requires not only the ability to generate feasible grasps, but also to ensure stability under multi-contact interactions, particularly in dual-arm
settings. While recent advances in data-driven grasp synthesis have shown promising results for
single-arm grasping, extending these methods to dual-arm scenarios introduces additional challenges such as coordinated contact placement, force balance, and collision avoidance. This thesis
addresses these challenges through a unified progression from generative modeling to analytical understanding, which are subsequently combined into an end-to-end framework for dual-arm grasp
generation.
We begin with Constrained GraspDiffusion Fields (CGDF), a diffusion-based generativemodelfor
6-DoF grasp synthesis on complex objects. CGDF introduces a part-guided diffusionmechanism that
enables sample-efficient generation of grasps constrained to specific regions of an object, allowing
dense and targeted grasp proposals even on large and geometrically intricate shapes. This provides
a strong foundation for generating diverse and region-aware grasp candidates.
Building on this, we recognize that dual-arm grasping requires a principled notion of stability beyond independent single-arm grasps. To address this, we developed DG16M, a large-scale dataset
and an analytical framework for evaluating dual-arm grasp stability using an optimization-based
force-closure formulation. By explicitly modeling contact forces, friction constraints, and external
disturbances, this work provides a physically grounded metric for assessing grasp quality and enables the learning of stability-aware models.
Finally, we integrate these two perspectives-- generativemodeling and analytical evaluation into
DAGDiff, an end-to-end diffusion-based framework for directly generating dual-arm grasp pairs.
Unlike prior approaches that rely on heuristic pairing or region selection, DAGDiff formulates grasp
generation in the joint SE(3)×SE(3)space and incorporates guidance from stability and collisionaware classifiers during the diffusion process. This allows the model to generate dual-arm grasps
that are not only diverse, but alsoforce-closure stable and collision-free directlyfrom raw input point
clouds.
Together, these contributions establish a principled pipeline for dual-arm grasp synthesis: from
region-aware generative modeling, to analytical stability evaluation, to fully integrated end-to-end
learning. This work advances the state of robotic manipulation by enabling reliable, generalizable,
and physically grounded dual-arm grasping on complex real-world objects.