Abstract

Machine Learning models increasingly face data integrity challenges due to the use of large-scale training datasets drawn from the Internet. We study what model developers can do if they detect that some data was manipulated or incorrect. Such manipulated data can cause adverse effects like vulnerability to backdoored samples, systematic biases, and in general, reduced accuracy on certain input domains. Machine Unlearning, traditionally studied for handling user data-deletion requests to provide privacy, can address these by allowing post-hoc deletion of affected training data from a learned model. Achieving perfect unlearning is computationally expensive; consequently, prior works have proposed inexact unlearning algorithms to solve this approximately as well as evaluation methods to test the effectiveness of these algorithms. In this thesis, we first outline some necessary criteria for evaluation methods and show no existing evaluation satisfies them all. Then, we design a stronger black-box evaluation method called the Interclass Confusion (IC) test which adversarially manipulates data during training to detect the insufficiency of unlearning procedures. We also propose two analytically motivated baseline methods (EU-k and CF-k) which outperform several popular inexact unlearning methods. We demonstrate how adversarial evaluation strategies can help in analyzing various unlearning phenomena which can guide the development of stronger unlearning algorithms. Next, we study the practical constraint that model developers may not know all manipulated training samples. Often, only a small, representative subset of the affected data is flagged. We formalize �Corrective Machine Unlearning� as the problem of mitigating the impact of data affected by unknown manipulations on a trained model, possibly knowing only a subset of impacted samples. We demonstrate that the problem of corrective unlearning has significantly different requirements from traditional privacyoriented unlearning. We find most existing unlearning methods, including the gold-standard retrainingfrom-scratch, require most of the manipulated data to be identified for effective corrective unlearning. However, one approach, SSD, achieves limited success in unlearning adverse effects with just a small portion of the manipulated samples, showing the tractability of this setting. We hope our work spurs research towards developing better methods for corrective unlearning. Finally, we demonstrate the use of unlearning in reducing the risk of Large Language Models assisting malicious use in the creation of bioweapons and cyberattacks. Adaptations of existing state-of-the-art unlearning techniques fail on this task, probably due to complexities introduced by not having access to training data that leads to such capabilities. We discuss Contrastive Unlearning Tuning (CUT), a Representation Engineering based unlearning method that steers models towards novice behaviour on potentially harmful dual-use knowledge, while retaining general model capabilities. We design a probing evaluation which shows CUT succeeds in removing this knowledge even from the internal layer representations of LLMs. Overall, this thesis attempts to extend the frontiers of unlearning from user-privacy applications to debiasing, denoising, removing backdoors, and removing harmful dual-use capabilities. We highlight the shortcomings of privacy-oriented unlearning methods and formulations in achieving these goals. We hope our work offers practitioners a new strategy to handle challenges arising from web-scale training, and post-training line of defense towards ensuring AI Safety.

Abstract

Knowledge graphs (KG), which are structures representing information corresponding to entities/topics and their inter-connections, have been playing a crucial role in leveraging information on the web for several downstream tasks such as text-generation, classification, etc. Hence, it becomes vital to construct and maintain such knowledge graphs. There are some previous efforts in populating such KGs and generating relevant entity/node embeddings for this task. However, these methods typically do not focus on analyzing entity-specific content exclusively, but rely on transformational techniques on a fixed collection of documents with certain entities. We define an approach to populate such KGs by utilizing entity-specific content on the web, for generating category-specific entity embeddings. We empirically prove our approach�s effectiveness, by utilizing it for a downstream task of Notability detection, associated with one of the most popular and important Knowledge Graphs - the Wikipedia platform. Wikipedia is a highly essential platform because of its informative, dynamic, and easily accessible nature. The rate of new content being uploaded to it is very high, which makes it essential to moderate this uploaded content. To ensure that only important and relevant content is uploaded to Wikipedia, its editors define a specific set of "Notability" guidelines. These guidelines indicate whether a particular title warrants its own Wikipedia article. So far notability is enforced by humans, which makes scalability an issue, and there has been no significant work on automating notability detection across diverse categories. We work on this problem of creating an automated system to detect the notability of different types of articles/pages, for a vast set of categories. It is not a trivial task to define a fixed procedure for determining the Notability of Wikipedia pages, as there are different types of pages in Wikipedia, in the way they correlate with the various categories in which they exist. For a given Wikipedia category, articles/pages that are simple category instances co-exist with pages that are associated with the category in a non-trivial manner. It is essential to distinguish Wikipedia pages based on this fundamental difference, to gauge the notability of the page accordingly, as the parameters to look for performing this Notability test vary in each case. We divide this problem into two components, based on the nature of an article�s title. We define two types of article titles - Simple titles and Complex titles. Simple titles correspond to simple category instances/named entities for a given category. For example, "Virat Kohli" is a simple title of the category "Cricket Players". Complex titles correspond to article titles that have complex dependencies with their category. For example, an article titled "Wake Island" might be present in the "Birds" category, because of its association with the category, but not because it represents an instance of a bird. This distinction helps us define the categories to analyze for generating category-specific embeddings. Articles with simple titles are further divided into two classes, the "Abstract" class and the "Generic" class, based on whether they represent abstract concepts (such as Temperature / Pressure) or not, respectively, as the process for notability detection is to be followed differently in each case. We construct a dataset with notable and non-notable samples, for 9 categories belonging to the Generic class and 5 categories belonging to the Abstract class. On the other hand, for articles with complex titles, another dataset is constructed for the 9 categories of the Generic class, as defining complex titles for conceptual entities in the Abstract class is non-trivial. We further design a generalizable mechanism to differentiate between simple and complex titles. Initially, we specifically worked on designing a notability detection system for articles with simple titles. This system is based on web-based entity features and their text-based salience encodings. We further incorporate neural networks and BERT encodings (transformer encoder) to perform binary classification. For validating our system�s performance in this task, we utilize accuracy metrics, correlation analysis, ablation study, and prediction confidence on popular Wikipedia pages. Our system outperforms machine learning-based classifier approaches and existing handcrafted entity salience detection algorithms. Further, we define a system to detect notability specifically for articles with complex titles. This system is primarily defined on the basis of web-based features and the salience of a title in its web-based documents� text. We train a Graph neural network (GNN) that generates attention-enhanced encodings for classification, with syntactic and semantic document graphs as inputs. We evaluated this system similar to the above syst

Abstract

The aging of the brain is a complex process shaped by a combination of genetic factors and environmental influences, exhibiting variations from one population to another. This thesis in- vestigates normative population-specific structural changes in the brain and explores variations in aging-related changes across different populations. The study gathers data from diverse groups, constructs individual models, and compares them through a thoughtfully designed framework. This thesis proposes as a comprehensive pipeline covering data collection, modeling, and the creation of an analysis framework. Finally, it offers an illustrative cross-population analysis, shedding light on the comparative aspects of brain aging. In our study, the Indian population is considered as the reference, and an effort is made to ad- dress gaps within this population through the creation of a population-specific database, an atlas, and an aging model to facilitate the study. Due to the challenges in data collection, we adopted a cross-sectional approach. A cross-sectional brain image database is meticulously curated for In- dian population. A sub-cortical structural atlas is created for the young population, enabling us to establish reference structural segmentation map for the Indian population. Age-specific, gender balanced, and high-resolution scans collected to create the first Indian brain aging model. Choosing cross-sectional data collection made sense because data from other populations were also mostly collected in a cross-sectional manner. Using the in-house database for Indian population and pub- licly available datasets for other populations, our inter-population analysis compares aging trends across Indian, Caucasian, Chinese, and Japanese populations. Developing an aging model from cross-sectional data presents challenges in distinguishing between cross-sectional variations and normative trends. In response, we proposed a method specifically tailored for cross-sectional data. We present a unique metric within our comprehensive aging comparison framework to differentiate between temporal and global anatomical variations across populations. This thesis has detailed a comprehensive process to compare the aspects of healthy aging across these diverse groups, ultimately concluding with a pilot study across four different populations. This framework can be readily adapted to study various research problems, exploring changes associated with different populations while considering factors beyond ethnicity, such as lifestyle, education, socio-economic factors, etc. Similar analysis frameworks and studies with multiple modalities and larger sample sizes will contribute to deriving more conclusive results.

Abstract

Factual inconsistencies in text, which include a range of errors from minor inaccuracies to substantial distortions, present a significant challenge in the realm of information dissemination. These inconsistencies, whether unintentional or the result of deliberate misinformation, can lead to a skewed understanding and flawed decision-making. In the context of the vast and complex landscape of digital data, the limitation of traditional verification methods becomes evident, highlighting the need for more advanced solutions. In this thesis, we present and explore three critical problems in this domain, each addressing a unique aspect of content credibility and factual consistency. The first problem we tackled in this thesis is the detection of hostility in online content, specifically focusing on Hindi tweets. Our approach categorizes these tweets into distinct hostile classes: hateful, offensive, defamatory, or fake. We employed pretrained Transformer-based models, particularly IndicBERT, which is adept at processing Hindi text due to its training on a vast corpus of Indian languages. The architecture of our model effectively utilizes information from emojis and hashtags, in addition to the natural language text. A significant enhancement in performance was achieved through Task Adaptive Pretraining (TAPT), leading to increases of 1.35% and 1.40% in binary hostility detection, and improvements of 4.06% and 1.05% in macro and weighted F1 metrics, respectively, for fine-grained classifications. Notably, our system, under the team name �iREL IIIT�, achieved first place in the �Hostile Post Detection in Hindi� shared task at the CONSTRAINT-2021 workshop. The second problem addressed in this thesis delves into the realm of automated fact extraction and verification, a pressing challenge in the digital landscape rife with misinformation. Central to our approach is the innovative Fact Extraction and Verification (FEVER) project, which assesses the veracity of claims against a comprehensive body of evidence from Wikipedia. Our contributions include the development of a specialized retrieval model tailored for the FEVER dataset, a strategic approach to sentence selection for optimal evidence gathering, and an exploration of advanced natural language inference (NLI) models, particularly state-of-the-art transformer models. By integrating these components, we not only refine the process of recognizing textual entailment but also significantly enhance the accuracy and efficiency of automated fact-checking. In tackling the third problem of this thesis, we address the critical issue of detecting and explaining factual inconsistencies in text, a significant challenge in the era of advanced Transformer-based natural language generation models. These models, while adept in tasks like summarization and translation, often struggle with producing hallucinatory and inconsistent content. Our approach introduces the novel Factual Inconsistency Classification with Explanations (FICLE) method. This technique involves a detailed analysis of sentence pairs to identify inconsistency types and provide comprehensive explanations, including inconsistent fact triples, context spans, and entity types. Central to our approach is the creation of the FICLE dataset, extensively annotated to cover a range of inconsistency types and their explanations. Utilizing this dataset, we developed a pipeline comprising four neural models, each designed to focus on specific facets of inconsistency detection and explanation. These models utilize various Transformer-based NLU and NLG architectures, with DeBERTa showing notable effectiveness across most sub-tasks. Our results underscore the effectiveness of this approach, demonstrating high performance in inconsistency type classification and entity-type prediction, with weighted F1 scores of around 87% and 86%, respectively, for these tasks. The detection of context spans, while more challenging, achieved an Intersection over Union (IoU) of approximately 65%, indicating the nuanced complexity of this aspect. The contributions of this research are multi-fold: proposing a new problem in factual inconsistency detection, creating a novel dataset, and establishing a baseline pipeline with high performance in inconsistency-type classification and entity-type prediction.

Abstract

All forms of art embody imagination, and it is medium-specific. This research investigates how photography embodies imagination, for the photography method differs from other art forms in its unique way. A painter builds the image stroke by stroke with paints. In photography, the canvas is partially filled with what exists before the camera. The processes involved in creating images, whether by analogue techniques or digital, point to the fact that it is not entirely automatic. Technology automates some features but involves thought processes, and the photographer makes precise decisions about it. There is some kind of 'occurring' at play. This research aims to understand the form of photography irrespective of the various contexts in which it is performed by gaining insights from Indian and Western philosophical traditions. This study attempts to knit together a formal theory of Photography from the viewpoint of the theory of Aucitya developed by the eleventh-century Kashmiri poet and critic K?emendra. It takes a step toward creating a dialogue between ancient theories of Indian aesthetics and contemporary visual arts. There have been few attempts to study photography in contemporary visual arts from Aucitya's viewpoint. This study proposes a theoretical framework for Photography based on Aucitya. It identifies a set of nine photographic determinants to explore a formal structure. The study claims that these photographic determinants allow one to break away from the closed movement of the default linear perspective embedded within the camera's architecture. By analysing various analogue and digital photographs, the research finds that these determinants become the impetus to fix a photographic frame, and the possible meanings ride on these determinants. Further, this study locates Aucitya in the works of artist-photographer Jyoti Bhatt, who is widely recognized as a painter, printmaker, and photographer. The interaction of various twodimensional visual media in his practice is interesting. This research attempts to show what photographic principles guide his practice.

Abstract

Intrusion Detection Systems (IDS) and Intrusion Prevention Systems (IPS) protect computer networks against unauthorised access and malicious traffic. False positives, where legitimate network traffic is incorrectly flagged as malicious, compromise the efficiency of these systems. False positives waste valuable resources and reduce the overall reliability of IDS/IPS alerts, leading to potential security risks. Our work aims to address the issue of false positive detection in IDS/IPS by utilising the capabilities of honeypots. Honeypots are deceptive systems designed to mimic legitimate services and attract potential attackers, providing a controlled environment to analyse their behaviour without posing any risk to the network infrastructure. This thesis explores the evolution of honeypots, from their beginning as passive research tools to their modern applications as refined deception mechanisms. The diverse types of honeypots are examined, indicating their potential to lure attackers and provide valuable insights into malicious activities without risking network assets. Next, the research analyses an in-depth survey of existing false positive mitigation techniques in IDS/IPS, analysing their strengths and limitations. Next, various honeypots are studied to identify their potential to effectively capture and differentiate legitimate and malicious traffic. The thesis proposes a hybrid approach that integrates the information gathered from honeypots with the output of the IDS/IPS, enabling more intelligent and precise decision-making for false positive detection. A comprehensive experimental setup is designed, simulating real-world network scenarios to evaluate the effectiveness of the proposed solution. The CSE-CIC-2018 dataset, a widely used and realistic cybersecurity dataset, is employed for experimentation to assess the effectiveness of our approach. Quantitative metrics such as true positive rate, false positive rate, and accuracy are measured to gauge the improvement achieved by the hybrid system. When honeypots are incorporated into the IDS/IPS framework, the results significantly reduce false positives. Furthermore, the study reveals insights into attackers� behaviour, aiding in developing more robust security policies and threat intelligence. In conclusion, this thesis presents a novel approach using honeypots to mitigate false positive detections in IDS/IPS. The integration of honeypots supports the network to make informed decisions, improving the overall efficiency and reliability of intrusion detection and prevention systems. By reducing false positives, networks can allocate resources more effectively and respond to genuine security threats.

Abstract

Remote labs are a groundbreaking development in the education industry, providing students with access to laboratory education anytime, anywhere in the world. This has proved to be particularly important during the recent COVID-19 pandemic, where remote access has become essential. However, most remote labs are costly and need more flexibility for institutions to replicate experiments. This becomes a significant concern when trying to scale up to accommodate more students in developing countries. The work presented in this thesis primarily revolves around building hardware for a proposed end-toend remote lab system (RLabs). This is presented in two parts: one is retrofitting Internet of Things (IoT) components to laboratory experiments, and the second is adopting Computer-Vision (CV) techniques in a remote lab experiment. The RLabs platform includes two use case experiments: Vanishing Rod and Focal Length. The hardware experiments are built at low cost by retrofitting IoT components. The proposed solution is qualitatively evaluated against seven non-functional attributes - affordability, portability, scalability, compatibility, maintainability, usability, and universality. Finally, user feedback was collected from a group of students, and the scores indicate a positive response to the student�s learning and the platform�s usability. Additionally, CV techniques are explored and used for a use case experiment of Conservation of Mechanical Energy to improve the accuracy and reduce the dependence on sensors used for remote experimentation. These CV techniques track the velocity of an object in the experiment, which is otherwise found out from infrared (IR) sensors. Later, velocities calculated from the CV technique are compared with the traditional IR sensors. The calculated velocities are further made accurate by employing linear regression.

Abstract

Two critical challenges arise in the interconnected realm of online platforms: the need to ensure equitable representation for entities and the emphasis on identifying deceptive practices. This work aims to address both by introducing CAFIN to decrease the disparity in the representations of GNNs, and by detecting and studying anomalies on the Google Play Store reviewer network for fraud prevention, moving in the direction of reshaping digital ecosystems to be both fair and trustworthy. Unsupervised Representation Learning on graphs is gaining traction due to the increasing abundance of unlabelled network data and the compactness, richness, and usefulness of the representations generated. In this context, the need to consider fairness and bias constraints while generating the representations has been well-motivated and studied to some extent in prior works. One major limitation of most of the prior works in this setting is that they do not aim to address the bias generated due to connectivity patterns in the graphs, such as varied node centrality, which leads to a disproportionate performance across nodes. In our work, we aim to address this issue of mitigating bias due to inherent graph structure in an unsupervised setting. To this end, we propose CAFIN, a centrality-aware fairness-inducing framework that leverages the structural information of graphs to tune the representations generated by existing frameworks. We deploy it on GraphSAGE (a popular framework in this domain) and showcase its efficacy on two downstream tasks - Node Classification and Link Prediction. Empirically, CAFIN consistently reduces the performance disparity across popular datasets (varying from 18% to 80% reduction in performance disparity) from various domains while incurring only a minimal cost of fairness. Google Play Store�s policy forbids the use of incentivized installs, ratings, and reviews to manipulate the placement of apps. However, there still exist apps that incentivize installs for other apps on the platform. To understand how install-incentivizing apps affect users, we examine their ecosystem through a socio-technical lens and perform a mixed-methods analysis of their reviews and permissions. Our dataset contains 319K reviews collected daily over five months from 60 such apps that cumulatively account for over 160.5M installs. We perform qualitative analysis of the reviews to reveal various types of dark patterns that developers incorporate in install-incentivizing apps, highlighting their normative concerns at both user and platform levels. Permissions requested by these apps validate our discovery of dark patterns, with over 92% apps accessing sensitive user information. We find evidence of fraudulent reviews on install-incentivizing apps, following which we model them as an edge stream in a dynamic bipartite graph of apps and reviewers. Our proposed reconfiguration of a state-of-the-art microcluster anomaly detection algorithm yields promising preliminary results in detecting this fraud. We discover highly significant lockstep behaviors exhibited by reviews that aim to boost the overall rating of an install-incentivizing app. Upon evaluating the 50 most suspicious clusters of boosting reviews detected by the algorithm, we find (i) near-identical pairs of reviews across 94% (47 clusters), and (ii) over 35% (1,687 of 4,717 reviews) present in the same form near-identical pairs within their cluster. We also discuss how fraud is intertwined with labor and poses a threat to the trust and transparency of Google Play

Abstract

Machine Learning (ML), particularly deep learning, has seen vast advancements, leading to the rise of Machine Learning-Enabled Systems (MLS). However, numerous software engineering challenges persist in propelling these MLS into production, largely due to various run-time uncertainties that impact the overall Quality of Service (QoS). These uncertainties emanate from ML models, software components, and environmental factors. Self-adaptation techniques present potential in managing runtime uncertainties, but their application in MLS remains largely unexplored. As a solution, this thesis proposes Machine Learning Model Balancer, a novel concept focusing on managing uncertainties related to ML models by using multiple models in runtime. Subsequently, the thesis introduces AdaMLS, an novel approach that leverages the Machine Learning Model Balancer concept for continuous adaptation. AdaMLS extends the traditional MAPE-K loop, employing lightweight unsupervised learning for dynamic model switching, thereby ensuring consistent QoS in dynamic environments. The effectiveness of AdaMLS is demonstrated through an object detection use case, showcasing its ability to effectively mitigate run-time uncertainties and surpass both naive approaches and standalone models in terms of QoS. We further developed SWITCH, an exemplar to demonstrate the practical application of our research in self-adaptation of MLS. The discussion on SWITCH highlights its role as a tool designed to enhance self-adaptive capabilities in MLS through dynamic model switching in runtime. SWITCH is developed as to cater to a broad range of ML scenarios. It features advanced input handling, real-time data processing, and logging for adaptation metrics, supplemented with an interactive real-time dashboard for system observability. Through its architecture and user-friendly interface, SWITCH not only demonstrates adaptability and performance but also serves as a valuable platform for researchers, practitioners, and students to explore self-adaptation in MLS. Beyond the primary focus on ensuring optimal QoS, we also explore the application of the Machine Learning Model Balancer concept in two main areas. Firstly, the EcoMLS approach leverages this concept to enhance the sustainability of MLS. By optimally balancing energy consumption with model confidence through runtime ML model switching, EcoMLS marks a significant step towards sustainable, energyefficient ML solutions. Secondly, RelMLS approach to systems, where machine learning is deployed in streaming mode. Through experiments and implementation, this approach has shown promising results in object detection use case, adopting a software architecture-based solution to dynamically switch between models based on contextual reliability. This thesis encapsulates a comprehensive exploration of the Machine Learning Model Balancer concept, from its theoretical introduction to practical applications in diverse MLS scenarios. Through AdaMLS, SWITCH, EcoMLS, and RelMLS implementations, we demonstrate the potential of our approaches in enhancing both the QoS and sustainability of MLS. Our findings suggest that the judicious application of model switching and self-adaptation techniques can significantly mitigate run-time uncertainties, paving the way for more resilient, efficient, and adaptable MLS.

Abstract

Sexual objectification is a form of objectification where a person is reduced to his/her sexual functionalities. In general, women and minority people are primary victims of sexual objectification, although sexualization of men has been on the rise. Initial research believed sexual objectification to be a behavioral phenomenon. That is, a person is sexually objectified through thoughts or actions of the objectifier. For example, studies have found a strong correlation between the body-biased gaze (greater attention to sexual body parts), also termed the objectifying gaze, and the sexual objectification of women. Recent studies indicate that the manifestation of sexual objectification may extend to sub-conscious cognitive processes. There are fundamental differences in brain activity when processing human stimuli versus object stimuli. A recent study indicates that when objectified, sexualized female images but not sexualized male images are visually processed more locally, similar to objects. For decades, there have been concerns regarding the growing hyper-sexualization of women in popular media like music videos, video games, etc., leading to content analysis studies on the music videos in various countries across the globe. And the results show that female characters� bodies are overly emphasized, primarily through revealing clothes and close-up shots of their bodies or sexual body parts. Frequent exposure to these objectifying media can lead to the manifestation of the objectifying gaze in the viewer. For example, a recent study reported that after watching sexualized pop music videos, viewers gaze objectified female images more. That is, they spent more time fixating on the sexual body parts of the female characters. In this thesis, we explore different forms of gaze objectification of women and the priming effect of sexually objectifying music videos on the viewers� gaze when looking at female images in a South Asian context. The thesis comprises of two studies. In the first study, we explore gaze objectification of nonsexualized female images. We have used the eye-tracking mechanism to capture the gaze pattern and the local/global paradigm to understand the underlying visual process when looking at non-sexualized female images. In the subsequent study, we investigate the priming effect of camera-induced objectifying gaze, operationalized via sexualized music video (MV), when looking at target female images in traditional (saree, salwar) and western (shirt-pant, short-dress) attire. Our results show both genders� gaze-objectified female sexual body parts. Participants spent more time looking at female sexual body parts. And male (vs. female) images are processed more locally, similar to objects. The findings suggest that women are gaze-objectified even in non-sexualized attire. In the second study, we found that, when primed with sexualized music videos (vs. non-sexualized), female images are more gaze-objectified. Our studies provide concrete evidence supporting existing research, which emphasizes that women experience elevated levels of gaze objectification. Additionally, exposure to sexualized media is identified as a factor contributing to the heightened objectification of women. This highlights the connection between media exposure and a cultural context where women are more commonly objectified. Recognizing and addressing these dynamics becomes crucial in understanding and mitigating the impact of sexual objectification.