Aravinda Raman Jatavallabha

Large Language Models (LLMs) have shown remarkable capabilities in privacy-sensitive domains like healthcare and hiring. However, concerns around regurgitating Personally Identifiable Information (PII) and Protected Health Information (PHI) persist. This study evaluates GPT-3.5, GPT-4, and GPT-4 Turbo, examining their propensity to leak sensitive data and the effectiveness of prompt-induced sanitization techniques. Using synthetic datasets reflecting HIPAA and GDPR compliance requirements, the models were benchmarked across privacy leakage, anonymization accuracy, and contextual utility. Our results show that GPT-4 and GPT-4 Turbo significantly reduce sensitive data leakage while preserving output utility, providing actionable insights for deploying LLMs in regulated environments.

In the world of Large Language Modeling, incremental learning plays an important role in evolving data such as streaming text. We introduce an incremental learning approach for dynamic contextualized word embeddings in the setting of streaming data. We call the embeddings generated by our model as Incremental Dynamic Contextualized Word Embeddings (iDCWE). Our model introduces the incremental BERT (iBERT) (BERT stands for Bidirectional Encoder Representations from Transformers) to create a dynamic and incremental model to perform incremental training. Our model further captures the semantic drift of words using dynamic graphs. Our paper is the first in the line of research on (incremental) dynamic modeling of streaming text which we also refer to as Neural Dynamic Language Modeling. The performance of our model on the benchmark datasets is on par and even often outperforms the dynamic contextualized word embeddings which was the first paper to combine contextualization with dynamic word embeddings, to the best of our knowledge. Moreover, the compute time efficiency of our model is better than that of the aforementioned paper.

The utilization of Graph Neural Networks (GNN) in modeling real-world graph structures has shown promising results, making it a widely recognized method for extracting information from non- Euclidean data, such as social networks. For static networks, there are many different sophisticated GNN architectures and models; nevertheless, the development of comparable methods for dynamic graphs has been slow. Due to the dynamic nature of graphs, Dynamic Graph Neural Networks have recently attracted interest in various fields, especially social networks. Continuous Time Dynamic Graphs (CTDG) effectively incorporate temporal information, graph structure (topology), and node properties to record the continuous time progression of dynamic graphs. The computational and memory requirements for Dynamic Graph Neural Networks provide considerable difficulties. In order to address this, a novel deep learning methodology using a model called TDGNN (Temporal Dynamic Graph Neural Network) has been developed to effectively model dynamic graphs while including temporal data, graph structure, and node attributes. The model suggests an approach that has been shown to perform better than baseline methods when tested on benchmark datasets.

The continuous growth of mobile traffic and limited spectrum resources limits the capacity and data rate. Heterogeneous Networks (HetNet) is a solution with multiple radio interfaces in smartphones to realize such demands. Simultaneous data transfer on Long Term Evolution (LTE) and WiFi has gained attention for data offloading in 5G HetNet. Maintaining the average throughput and minimum delay for LTE users is still a challenge in data offloading owing to the mobility and load in the network. This study explores the benefits of Software-Defined Networking (SDN) based multipath for data offloading schemes for LTE-WiFi integrated networks to maintain the user's average throughput based on channel quality classification. We classify future link qualities using deep learning models such as Long Short-Term Memory Networks (LSTM) and Bidirectional Long Short-Term Memory Networks (BLSTM). The received signal strength indicator (RSSI) and packet data rate (PDR) are parameters used in BLSTM. The results of the prediction were compared with those of state-of-the-art methods. We obtained a 2.1% better prediction than the state-of-the-art methods. The predicted results were used to offload the data using LTE and WiFi. The performance of HetNet was compared with the state-of-the-art method for average throughput, and with the proposed method, a 6.29% improvement was observed.

This paper introduces a feature distillation framework that aims to learn fairer representations without significantly sacrificing task performance. We propose a Maximum Mean Discrepancy (MMD)-based loss to distill information from an unfair teacher network while encouraging feature invariance across protected attributes such as race or gender. Our method demonstrates marked reduction in disparity measures while maintaining competitive accuracy on standard computer vision benchmarks.

This paper presents a hierarchical transformer-based framework that jointly models textual and visual modalities for detecting derailment in multimodal discussions. The proposed system integrates BERT-based text encoding with Faster R-CNN-derived visual features, achieving 71.0% accuracy and 78.3% AUC, outperforming text-only baselines by 6.1%. Our approach demonstrates the importance of multimodal cues in understanding conversation dynamics and early detection of potential derailments.

GraphNeural Networks (GNNs)have emerged as apotent framework for graph-structured recommendation tasks. Incorporating contrastive learning with GNNs has recently demon strated remarkable efficacy in addressing challenges posed by data sparsity, thanks to in novative data augmentation strategies. However, many existing methods employ stochastic perturbations (e.g., node or edge modifications) or heuristic approaches (e.g., clustering based augmentations) to generate contrastive views, which may distort semantic integrity or amplify noise. We introduce LightGCL, a novel and streamlined graph contrastive learn ing model to address these limitations. LightGCL utilizes Singular Value Decomposition (SVD) to achieve robust augmentation, facilitating structural refinement and global collab orative relation modeling without manual augmentation strategies. Extensive experiments on benchmark datasets showcase its substantial performance enhancements over state-of the-art methods. Further analysis highlights the model's resilience to challenges like data sparsity and bias related to item popularity.

This case study delves into the ethical ramifications of incidents involving Tesla's Autopilot, emphasizing Tesla Motors' moral responsibility. Using a seven-step ethical decision-making process, we examine user behavior, system constraints, and regulatory implications. The analysis offers insights into ethical considerations in evolving technological landscapes and proposes a framework for evaluating autonomous system deployments in safety-critical applications.

This research investigates flight delay trends by examining factors such as departure time, airline, and weather conditions. We employ time-series models including LSTM, Hybrid LSTM, and Bi-LSTM, comparing them with baseline regression models. Our approach focuses on identifying influential features in delay prediction, potentially informing flight planning strategies. The study demonstrates the effectiveness of deep learning approaches in capturing complex temporal patterns in aviation operations.

This study addresses the critical need for early diabetes detection using machine learning algorithms. We compare K-Nearest Neighbor, Random Forest, and Artificial Neural Network approaches, incorporating comprehensive preprocessing and feature engineering strategies. The Random Forest classifier achieved the highest accuracy of 87.89%, significantly outperforming traditional diagnostic methods and demonstrating the potential for ML-assisted medical decision-making.

This research evaluates deep learning models for automated bone age assessment, comparing pre-trained architectures including VGG-16, InceptionV3, XceptionNet, and MobileNet. Our study focuses on pediatric X-ray images, developing a system that can accurately determine skeletal maturity without expert intervention. The results demonstrate the potential for AI-assisted growth evaluation in clinical settings, with particular emphasis on reducing assessment time while maintaining accuracy.