Federated Learning (FL) has made it possible to learn from data that would’ve otherwise not been possible due to privacy and security restrictions. FL relies on each client’s honest participation and contribution. The existence of free-rider attackers may undermine the FL process allowing the free-riders to enjoy the contributions of the honest clients without any of their own. In FL algorithms such as Federated Averaging (FedAvg), the goal of the free-rider is to obtain the global model. However, algorithms such as Federated Model Distillation (FedMD) do not have a global model. This enables clients to train models with unique architectures. Here, collaborative learning is done by sharing prediction logits on a common public dataset aggregated by a central server. In this article, we propose a free-riding attack specific to this scenario. Here, the goal of the free-rider is to steal the aggregated logits from the server. Our proposed free-rider attack exploits the model heterogeneity property of FedMD and utilizes fake models to create the prediction logits. Furthermore, we improve upon FedMD and develop a KL-divergence-based detection mechanism to defend against such fake model attacks. Our experiments show that our mechanism can remove such free-riders at the very start of the FL process. Additionally, we also provide theoretical justification for the covertness of the fake model attack and the effectiveness of our detection mechanism.

Recent breakthroughs in Large Language Models (LLM), comprised of billions of parameters, have achieved the ability to unveil exceptional insight into a wide range of Natural Language Processing (NLP) tasks. The onus of the performance of these models lies in the sophistication and completeness of the input prompt. Minimizing the enhancement cycles of prompt with improvised keywords becomes critically important as it directly affects the time to market and cost of the developing solution. However, this process inevitably has a trade-off between the learning curve/proficiency of the user and completeness of the prompt, as generating such a solutions is an incremental process. In this paper, we have designed a novel solution and implemented it in the form of a plugin for Visual Studio Code IDE, which can optimize this trade-off, by learning the underlying prompt intent to enhance with keywords. This will tend to align with developers' collection of semantics while developing a secure code, ensuring parameter and local variable names, return expressions, simple pre and post-conditions. and basic control and data flow are met.

Autonomous agents equipped with Large Language Models (LLMs) are rapidly gaining prominence as a revolutionary technology within the realm of Software Engineering. These intelligent and autonomous systems demonstrate the capacity to perform tasks and make independent decisions, leveraging their intrinsic reasoning and decision-making abilities. This paper delves into the current state of autonomous agents, their capabilities, challenges, and opportunities in Software Engineering practices. By employing different prompts (with or without context), we conclude the advantages of contextrich prompts for autonomous agents. Prompts with context enhance user requirement understanding, avoiding irrelevant details that could hinder task comprehension and degrade model performance, particularly when dealing with complex frameworks such as Spring Boot, Django, Flask, etc. This exploration is conducted using Auto-GPT (v0.3.0), an open-source application powered by GPT-3.5 and GPT-4 which intelligently connects the “thoughts” of Large Language Models (LLMs) to independently accomplish the assigned goals or tasks.

Deploying applications on hybrid clouds with computational artifacts distributed over public backends and private edges involve several constraints. Designing such deployment requires application architects to solve several challenges, spanning over hard regulatory policy constraints as well as business policy constraints such as enablement of privacy by on-prem processing of data to the extent the business wants, backend support of privacy enabling technologies (PET), sustainability in terms of green energy utilization, latency sensitivity of the application. In this paper, we propose to optimize hybrid cloud application architectures, while taking all those factors into consideration, and empirically demonstrate the effectiveness of our approach. To the best of our knowledge, this work is the first of its kind.

Autism Spectrum Disorder (ASD) is a universal and often lifelong neuro-developmental disorder. Individuals with ASD often present comorbidities such as epilepsy, depression, and anxiety. In the United States, in 2014, 1 out of 68 people was affected by autism, but worldwide, the number of affected people drops to 1 in 160. This disparity is primarily due to underdiagnosis and unreported cases in resource-constrained environments. Wiggins et al. 1 found that, in the US, children of color are under-identified with ASD. Missing a diagnosis is not without consequences; approximately 26% of adults with ASD are under-employed, and are under-enrolled in higher education.

Any given AI system cannot be accepted unless its trustworthiness is proven. An important characteristic of a trustworthy AI system is the absence of algorithmic bias. 'Individual discrimination' exists when a given individual different from another only in 'protected attributes' (e.g., age, gender, race, etc.) receives a different decision outcome from a given machine learning (ML) model as compared to the other individual. The current work addresses the problem of detecting the presence of individual discrimination in given ML models. Detection of individual discrimination is test-intensive in a black-box setting, which is not feasible for non-trivial systems. We propose a methodology for auto-generation of test inputs, for the task of detecting individual discrimination. Our approach combines two well-established techniques - symbolic execution and local explainability for effective test case generation. We empirically show that our approach to generate test cases is very effective as compared to the best-known benchmark systems that we examine.

For document level sentiment analysis (SA), Subjectivity Extraction, ie., extracting the relevant subjective portions of the text that cover the overall sentiment expressed in the document, is an important step. Subjectivity Extraction, however, is a hard problem for systems, as it demands a great deal of world knowledge and reasoning. Humans, on the other hand, are good at extracting relevant subjective summaries from an opinionated document (say, a movie review), while inferring the sentiment expressed in it. This capability is manifested in their eye-movement behavior while reading: words pertaining to the subjective summary of the text attract a lot more attention in the form of gaze-fixations and/or saccadic patterns. We propose a multi-task deep neural framework for document level sentiment analysis that learns to predict the overall sentiment expressed in the given input document, by simultaneously learning to predict human gaze behavior and auxiliary linguistic tasks like part-of-speech and syntactic properties of words in the document. For this, a multi-task learning algorithm based on multi-layer shared LSTM augmented with task specific classifiers is proposed. With this composite multi-task network, we obtain performance competitive with or better than state-of-the-art approaches in SA. Moreover, the availability of gaze predictions as an auxiliary output helps interpret the system better; for instance, gaze predictions reveal that the system indeed performs subjectivity extraction better, which accounts for improvement in document level sentiment analysis performance.

Measuring reading effort is useful for practical purposes such as designing learning material and personalizing text comprehension environment. We propose a quantification of reading effort by measuring the complexity of eye-movement patterns of readers. We call the measure Scanpath Complexity. Scanpath complexity is modeled as a function of various properties of gaze fixations and saccades- the basic parameters of eye movement behavior. We demonstrate the effectiveness of our scanpath complexity measure by showing that its correlation with different measures of lexical and syntactic complexity as well as standard readability metrics is better than popular baseline measures based on fixation alone.

Cognitive NLP systems-i.e., NLP systems that make use of behavioral data -augment traditional text-based features with cognitive features extracted from eye-movement patterns, EEG signals, brain-imaging etc..Such extraction of features is typically manual.We contend that manual extraction of features may not be the best way to tackle text subtleties that characteristically prevail in complex classification tasks like sentiment analysis and sarcasm detection, and that even the extraction and choice of features should be delegated to the learning system.We introduce a framework to automatically extract cognitive features from the eye-movement / gaze data of human readers reading the text and use them as features along with textual features for the tasks of sentiment polarity and sarcasm detection.Our proposed framework is based on Convolutional Neural Network (CNN).The CNN learns features from both gaze and text and uses them to classify the input text.We test our technique on published sentiment and sarcasm labeled datasets, enriched with gaze information, to show that using a combination of automatically learned text and gaze features often yields better classification performance over (i) CNN based systems that rely on text input alone and (ii) existing systems that rely on handcrafted gaze and textual features.

The sentiment aggregation problem accounts for analyzing the sentiment of a user towards various aspects/features of a product, and meaningfully assimilating the pragmatic significance of these features/aspects from an opinionated text. The current paper addresses the sentiment aggregation problem, by assigning weights to each aspect appearing in the user-generated content, that are proportionate to the strategic importance of the aspect in the pragmatic domain. The novelty of this paper is in computing the pragmatic significance (weight) of each aspect, using graph centrality measures (applied on domain specific ontology-graphs extracted from ConceptNet), and deeply ingraining these weights while aggregating the sentiments from opinionated text. We experiment over multiple real-life product review data. Our system consistently outperforms the state of the art - by as much as a F-score of 20.39% in one case.

Gaze based navigation with digital screens offer a hands-free and touchless interaction, which is often useful in providing a hygienic interaction experience in a public kiosk scenario. The goodness of such a navigation system depends not only on the accuracy of detecting the eye gaze but also on the ability to determine whether a user is interested in clicking a button or is just looking at the button. The time for which a user needs to gaze at a particular button before it is considered as a click action is called the dwell time. In this paper, we explore intelligent adjustment of dwell times, where mouse click events on the buttons of a given application are emulated with user gaze. A constant dwell-time for all buttons and for all users may not provide an efficient and intuitive interface. We thereby propose a model to dynamically adjust dwell-time values used to emulate user mouse click events, exploiting the user's experience with different portions of a given application. The adjustment happens at a per-user, per-button granularity, as a function of the user's (a) prior usage experience of the given button within the application and (b) Midas touch characteristics for the given button. We propose OptiDwell, inspired by the action-value method based solutions to the Multi-Armed Bandits problem, for dwell click time adaptation. We experiment OptiDwell using an interactive TV channel browsing interface application, constituting of a mix of text and image buttons, over 10 computer-savvy users generating over 9000 click tasks. We observe significant improvement of user comfort level over the sessions, quantified by (a) improved (reduced) dwell times and (b) reduced number of Midas touches in spite of faster dwell-clicks, as high as 10-fold reduction in the best case. Our work is useful for creating an interface, with accurate, fast and comfortable dwell-clicks for each interface element (e.g., buttons), and each user.

In this paper, we propose a novel mechanism for enriching the feature vector, for the task of sarcasm detection, with cognitive features extracted from eye-movement patterns of human readers.Sarcasm detection has been a challenging research problem, and its importance for NLP applications such as review summarization, dialog systems and sentiment analysis is well recognized.Sarcasm can often be traced to incongruity that becomes apparent as the full sentence unfolds.This presence of incongruity-implicit or explicit-affects the way readers eyes move through the text.We observe the difference in the behaviour of the eye, while reading sarcastic and non sarcastic sentences.Motivated by this observation, we augment traditional linguistic and stylistic features for sarcasm detection with the cognitive features obtained from readers eye movement data.We perform statistical classification using the enhanced feature set so obtained.The augmented cognitive features improve sarcasm detection by 3.7% (in terms of Fscore), over the performance of the best reported system.

This paper explores the social quality (goodness) of community structures formed across Twitter users, where social links within the structures are estimated based upon semantic properties of user-generated content (corpus). We examined the overlap of the community structures of the constructed graphs, and followership-based social communities, to find the social goodness of the links constructed. Unigram, bigram and LDA content models were empirically investigated for evaluation of effectiveness, as approximators of underlying social graphs, such that they maintain the community social property. Impact of content at varying granularities, for the purpose of predicting links while retaining the social community structures, was investigated. 100 discussion topics, spanning over 10 Twitter events, were used for experiments. The unigram language model performed the best, indicating strong similarity of word usage within deeply connected social communities. This observation agrees with the phenomenon of evolution of word usage behavior, that transform individuals belonging to the same community tending to choose the same words, made by [1], and raises a question on the literature that use, without validation, LDA for content-based social link prediction over other content models. Also, semantically finer-grained content was observed to be more effective compared to coarser-grained content.

Existing systems deliver high accuracy and F1-scores for detecting paraphrase and semantic similarity on traditional clean-text corpus. For instance, on the clean-text Microsoft Paraphrase benchmark database, the existing systems attain an accuracy as high as 0:8596. However, existing systems for detecting paraphrases and semantic similarity on user-generated short-text content on microblogs such as Twitter, comprising of noisy and ad hoc short-text, needs significant research attention. In this paper, we propose a machine learning based approach towards this. We propose a set of features that, although well-known in the NLP literature for solving other problems, have not been explored for detecting paraphrase or semantic similarity, on noisy user-generated short-text data such as Twitter. We apply support vector machine (SVM) based learning. We use the benchmark Twitter paraphrase data, released as a part of SemEval 2015, for experiments. Our system delivers a paraphrase detection F1-score of 0.717 and semantic similarity detection F1-score of 0.741, thereby significantly outperforming the existing systems, that deliver F1-scores of 0.696 and 0.724 for the two problems respectively. Our features also allow us to obtain a rank among the top-10, when trained on the Microsoft Paraphrase corpus and tested on the corresponding test data, thereby empirically establishing our approach as ubiquitous across the different paraphrase detection databases.

Sentiments expressed in user-generated short text and sentences are nuanced by subtleties at lexical, syntactic, semantic and pragmatic levels. To address this, we propose to augment traditional features used for sentiment analysis and sarcasm detection, with cognitive features derived from the eye-movement patterns of readers. Statistical classification using our enhanced feature set improves the performance (F-score) of polarity detection by a maximum of 3.7% and 9.3% on two datasets, over the systems that use only traditional features. We perform feature significance analysis, and experiment on a held-out dataset, showing that cognitive features indeed empower sentiment analyzers to handle complex constructs.

In this paper, we propose a supervised learning based model for ocular biometrics. Using Speeded-Up Robust Features (SURF) for detecting local features of the eye region, we create a local feature descriptor vector of each image. We cluster these feature vectors, representing an image as a normalized histogram of membership to various clusters, thereby creating a bag-of-visual-words model. We conduct a multiphase training, first performing a fast Multinomial Naïve Bayes learning, and subsequently using a pyramid-up topology to use the top k% results (based upon confidence scores) thus predicted and perform Dense SIFT for nearest neighbor matching. Contrary to traditional ocular biometric systems, our proposed approach does not rely highly accurate iris pattern segmentation, allowing less constrained image acquisition conditions such as from mobile devices. Our method identifies the individuals with an identification accuracy varying from 48.76% to 79.49%, across different lighting conditions and phone handset data sources, while testing on the given data.

We propose a geometrical method, applied over eye-specific features, to improve the accuracy of the art of eye-center localization. Our solution is built upon: (a) checking radially constrained gradient vectors, (b) adding weightage to iris specific features and (c) considering bi-directional image gradients to eliminate errors due to reflection on pupil. Our system outperforms the state of the art methods, when compared collectively across multiple benchmark databases, such as BioID and FERET. Our process is lightweight, robust and significantly fast: achieving 50-60 fps for eye center localization, using a single threaded approach on a 2.4 GHz CPU with no GPU. This makes it practicable for real-life applications.

Ocular biometrics in the visible spectrum has emerged as an area of significant research activity. In this paper, we propose two convolution-based models for verifying a pair of periocular images containing the iris, and compare the two approaches amongst each other as well as with a baseline model. In the first approach, we perform deep learning in an unsupervised manner using a stacked convolutional architecture, using external models learned a-priori on external facial and periocular data, on top of the baseline model applied on the provided data, and apply different score fusion models. In the second approach, we again use a stacked convolution architecture; but here, we learn the feature vector in a supervised manner. We obtain an AUROC of 0.946 and 0.981, and EER of 0.092 and 0.066, for the two models respectively. We further combine the two models, and observe the combined model to deliver the best performance in case the both the images arise from the same device type, but not necessarily so otherwise, obtaining a AUROC of 0.985 and EER of 0.057. Given the significant performance our methodology yields, our system can be used in real-life applications with minimal error.

Mobile browsing habits are characteristically different from browsing on traditional devices. Mobile users often look for information snippets instead of complete web pages. Also, mobile devices are often constrained in terms of resource availability, such as battery, data plan limits and network bandwidth. Under such constraints, partial-loading of a web page, by loading the most relevant content snippets early, can satisfy the user and consume less resources. Mobile content adaptation middleware has traditionally focused on user factors, such as user feedbacks and user context. We believe that the content creator, with complete knowledge of the importance of each item in the web page, is well-suited to guide the adaptation process. Combining user choice, and ratings assigned by the content creator to different web page elements (items), enables delivering the most relevant items in an ordered manner in response to a page request. We present ScoDA, a cooperative content adaptation middleware framework, under resource constraints. We evaluate the effectiveness of page loading using ScoDA, on simulated complex web pages as well as real web pages.

Here we study a variant of maximal clique enumeration problem by incorporating a minimum size criterion. We describe preprocessing techniques to reduce the graph size. This is of practical interest since enumerating maximal cliques is a computationally hard problem and the execution time increases rapidly with the input size. We discuss basics of an algorithm for enumerating large maximal cliques which exploits the constraint on minimum size of the desired maximal cliques. Social networks are prime examples of large sparse graphs where enumerating large maximal cliques is of interest. We present experimental results on the social network formed by the call detail records of one of the world's largest telecom service providers. Our results show that the preprocessing methods achieve significant reduction in the graph size. We also characterize the execution behaviour of our large maximal clique enumeration algorithm.