A comprehensive analysis was performed using the AIDS Clinical Trials Group 175 dataset to improve the accuracy of predicting AIDS disease progression. The primary objective was to integrate machine learning technique...
详细信息
Legal Outcome Prediction(LOP) is the process of predicting the possible outcome of a court case based on the contents in the case *** aim of this work is to perform a comparative analysis to assess the effectiveness o...
详细信息
Brain tumors using MRI images were a long-standing issue for mankind. Early disease detection and treatment leads to a better quality of but relying on human instincts alone is not enough. In recent times many advance...
详细信息
Chatbots powered by Large Language Model(LLM) can be manipulated by malicious prompts, generating harmful content and biased responses which would raise security concerns. Growing dependence on chatbots demands robust...
详细信息
ISBN:
(纸本)9798350369083
Chatbots powered by Large Language Model(LLM) can be manipulated by malicious prompts, generating harmful content and biased responses which would raise security concerns. Growing dependence on chatbots demands robust security for ethical development and user trust, which makes the work relevant in today's world. The motivation behind the work is to let the user have a safe experience with no negative responses being displayed while using the chatbot, which paved the way to arrive at the goal of developing a security filter that could be integrated into any LLM feature integrated application to mitigate the risk of having security vulnerabilities like prompt injection and jailbreaking, which could be achieved by converting malicious prompt into safer prompts by the method of eliminating negative sentiment phrases. The work focuses on building and implementing the security filters to popular in-production LLMs like Large Language Model Meta AI-2 (LLaMA2) and Generative Pre-trained Transformer - 3.5 turbo (GPT-3.5) to see how they handle against prompt injection and jailbreaking before and after the security filter being integrated. A large database of 200,000 prompts has been collected and preprocessed to train on a machine learning model using binary classification algorithm having 99.7% accuracy for classification of prompts into malicious or non-malicious and further checks are being done by breaking the prompt into smaller phrases and individually analyzing their compound sentiment score using Natural Language Toolkit (NLTK) Valence Aware Dictionary for Sentiment Reasoning (VADER) algorithm to detect and drop the negative sentiment phrases for the modification of the user prompt to eliminate the possibility of malicious prompt being passed to LLM. It is difficult to determine the sentiment of prompts in a detailed way and convert it into an efficient design that will perform well with models. Once this hurdle is overcome, chatbots will become even more reliable,
Combating zero-day attacks is essential in the age of ongoing cyber threats. For simulating and identifying these dangers, our research uses Long Short-Term Memory (LSTM) algorithms, which are skilled at collecting te...
详细信息
Telecare Medical Information System (TMIS) is a technology that has been on rise in the recent times with the onset of the COVID-19 pandemic. TMIS can be implemented in wireless body area networks (WBAN) in the form o...
详细信息
Quantum computing is a revolution in computational theory that uses the principles of quantum mechanics to perform computations that are beyond the capacity of classical computers. It works with the special features o...
详细信息
In the ever-evolving landscape of travel planning, the 'Trip Planner' application is highly recommended. Using Python, Tkinter, and SQLite managed through pandas, this tool enhances user interaction with techn...
详细信息
India is an agricultural country, supporting a significant portion of the population. Agriculture is a source of income for millions of Indians. But due to lack of knowledge and the state's diverse climate and soi...
详细信息
Optimizing Python code is essential for enhancing performance and efficiency. This project investigates the use of large pre-trained language models, specifically GPT (Generative Pre-trained Transformer), for Python c...
详细信息
暂无评论