Publications

1. The Alchemy of Thought: Understanding In-Context Learning Through Supervised Classification

   (Read Abstract)

   In-context learning (ICL) has become a prominent paradigm to rapidly customize LLMs to new tasks without fine-tuning. However, despite the empirical evidence of its usefulness, we still do not truly understand how ICL works. In this paper, we compare the behavior of in-context learning with supervised classifiers trained on ICL demonstrations to investigate three research questions: (1) Do LLMs with ICL behave similarly to classifiers trained on the same examples? (2) If so, which classifiers are closer, those based on gradient descent (GD) or those based on k-nearest neighbors (kNN)? (3) When they do not behave similarly, what conditions are associated with differences in behavior? Using text classification as a use case, with six datasets and three LLMs, we observe that LLMs behave similarly to these classifiers when the relevance of demonstrations is high. On average, ICL is closer to kNN than logistic regression, giving empirical evidence that the attention mechanism behaves more similarly to kNN than GD. However, when demonstration relevance is low, LLMs perform better than these classifiers, likely because LLMs can back off to their parametric memory, a luxury these classifiers do not have.

1. Finding a Wolf in Sheep's Clothing: Combating Adversarial Text-To-Image Prompts with Text Summarization

   (Read Abstract)

   Text-to-image models are vulnerable to the stepwise “Divide-and-Conquer Attack” (DACA) that utilize a large language model to obfuscate inappropriate content in prompts by wrapping sensitive text in a benign narrative. To mitigate stepwise DACA attacks, we propose a two-layer method involving text summarization followed by binary classification. We assembled the Adversarial Text-toImage Prompt (ATTIP) dataset (N = 940), which contained DACA-obfuscated and nonobfuscated prompts. From the ATTIP dataset, we created two summarized versions: one generated by a small encoder model and the other by a large language model. Then, we used an encoder classifier and a GPT-4o classifier to perform content moderation on the summarized and unsummarized prompts. When compared with a classifier that operated over the unsummarized data, our method improved F1 score performance by 31%. Further, the highest recorded F1 score achieved (98%) was produced by the encoder classifier on a summarized ATTIP variant. This study indicates that pre-classification text summarization can inoculate content detection models against stepwise DACA obfuscations.

1. Activity-Aware Data Rate Tuning in Wireless Body Area Networks

   (Read Abstract)

   This work proposes an Activity-Aware Data Rate Tuning (A2D) scheme for Wireless Body Area Network (WBAN), while considering the criticality of the physiological sensed data. We consider different physical activities of the patients and thereafter, compute their health criticality. Further, on the basis of the health criticality value, the data rate of these physiological sensors are tuned. Depending on the physical activity of a patient, the value sensed by the physiological sensors may change. Consequently, when a healthy person runs, a particular sensor value may be significantly high, even if it is normal, however, the same data reading may be critical for a person who is sitting or standing. Thus, a WBAN is required to be activity-aware in order to measure the correct criticality values. We implemented in a real hardware platform system to show the effectiveness of the proposed scheme. Experimental results show that the proposed scheme is capable of tuning the data rate of different physiological sensors, based on human activity and critical conditions, while ensuring more than 90% of packet delivery ratio in intra-BAN communication and 93% in inter-BAN communication.

1. A Computer Vision Framework for Partitioning of Image-Object Through Graph Theoretical Heuristic Approach

   (Read Abstract)

   The aim of this work is to develop a graph theoretical computer vision framework to partition shape of an image object into parts based on a heuristic approach such that the partitioning remains consistent with human perception. The proposed framework employs a special polygonal approximation scheme to represent a shape suitably in simpler graph form where each polygonal side represents a graph-edge. The shape-representative graph is explored to determine vertex-visibility graph by a simple algorithm presented in this paper. We have proposed a heuristic based iterative clique extraction strategy to decompose the shape-representative graph depending on its vertex-visibility graph. This proposed framework considers MPEG-7 shape data set for probing the acceptability of the proposed framework and according to our observation, the performance of the framework is comparable with existing schemes.