Maisha Binte Rashid

Ph.D. Candidate, Computer Science, Baylor University

Welcome! My name is Maisha and I am a Ph.D. candidate in Computer Science at Baylor University, where I am advised by Pablo Rivas. I am part of the Rivas Lab, specializing in multimodal machine learning and AI robustness. My research focuses on integrating text and image data to defend against adversarial attacks, with particular expertise in large-scale language models (LLMs), natural language processing, and deep learning.

As a Graduate Research Assistant, I have contributed to NSF-funded projects and enhanced AI safety for models like GPT-2 and Vision Transformer. My work includes developing the Adversarial Vulnerability Index (AVI), a novel metric for assessing risk in vision-language models through progressive stress testing.

Prior to my doctoral studies, I worked as a Programmer Analyst at Computer Ease Ltd. in Bangladesh, where I led development of B2B transaction systems and client dashboards, significantly improving processing efficiency and data fidelity.

Research Interests: Multimodal Machine Learning, AI Safety & Robustness, Adversarial Attacks & Defense, Vision-Language Models, Natural Language Processing, Deep Learning

Research & Experience

Graduate Research Assistant

Baylor University, Waco, TX

August 2021 – Present

Enhanced AI safety for models like GPT-2 and Vision Transformer, defending against adversarial attacks with integrated image and text data, boosting accuracy by 12%
Led multimodal deep learning project focusing on image and text data analysis for NSF-funded research
Developed the Adversarial Vulnerability Index (AVI) metric for vision-language models, using progressive stress testing from random noise to gradient-based attacks
Utilized Flamingo model to identify stolen car parts from comprehensive 400 GB dataset of Craigslist images and posts
Directed student engagement in graduate Natural Language Processing (NLP) course, advancing competencies in machine learning models and text analysis

Programmer Analyst

Computer Ease Ltd., Dhaka, Bangladesh

January 2020 – July 2021

Led development of 20 modules across 4 websites for B2B transactions using .NET and C#, reducing processing time by 30%
Developed two client dashboards for daily transaction monitoring using Angular and .NET Core, dropping error rates by 40%
Enhanced client operations and data fidelity through optimized system architecture

Publications

A Framework for Evaluating Vision-Language Model Safety: Building Trust in AI for Public Sector Applications

M. Rashid, P. Rivas

AAAI 2025 Workshop on AI for Social Impact: Bridging Innovations in Finance, Social Media, and Crime Prevention

Paper

AI Safety in Practice: Enhancing Adversarial Robustness in Multimodal Image Captioning

M. Rashid, P. Rivas

In Proceedings of ACM KDD 2024 (KDD ’24). ACM, New York, NY, USA

Paper

Leveraging OpenFlamingo for Multimodal Embedding Analysis of C2C Car Parts Data

M. Rashid, P. Rivas

World Congress in Computer Science, Computer Engineering & Applied Computing (Springer Nature Switzerland), July 22, 2024

Paper

Navigating the multimodal landscape: A review on integration of text and image data in machine learning architectures

M. Rashid, MS Rahaman, P. Rivas

Machine Learning and Knowledge Extraction Journal

Paper

Scoping Review on Image-Text Multimodal Machine Learning Models

M. Rashid, P. Rivas

International Conference on Computational Science and Computational Intelligence (CSCI), 2023

Paper

Autonomous deblurring images and information extraction from documents using CycleGAN and mask RCNN

O. Hoque, M. Rashid, T. Zawad

23rd International Conference on Computer and Information Technology (ICCIT) (IEEE), 2020

Paper

Targeted face recognition and alarm generation for security surveillance using single shot multibox detector (SSD)

T Jawad, B Maisha, S Nazmus

Int. J. Comput. Appl, 2019

Paper

Selected Projects

Adversarial Robustness in Vision Language Models

Developed multimodal machine learning model for image captioning using ViT and GPT-2:

Performed adversarial training to identify vulnerable modalities in the architecture
Strategically froze less vulnerable components to focus training on weaker modalities
Increased model performance from 74% to 78% through targeted robustness improvements

→ View Project

Analyzing Adversarial Vulnerability of Stable Diffusion

Comprehensive analysis of text-to-image generation model robustness:

Fine-tuned Stable Diffusion on COCO dataset for high-quality image generation
Applied gradient-based attacks (FGSM, PGD) to image latent components and text-based attacks to prompts
Measured robustness using LPIPS (perceptual change) and CLIP score (image-text alignment)

→ View Project

Adversarial Sensitivity Analysis Metric for VLMs

Novel framework for evaluating vision-language model vulnerabilities:

Conducted controlled modality stress test on CLIP by alternately freezing ViT and text encoders
Discovered vision modality showed 9% larger accuracy drop than text modality on Hateful Memes dataset
Defined Adversarial Vulnerability Index (AVI) for progressive evaluation from random noise to adversarial attacks

→ View Project

CyberPolice - Cyberbullying Detection System

Deep learning approach for detecting cyberbullying in social media:

Developed three models: LSTM, Bi-LSTM, and Bi-LSTM with Transformer for Facebook post analysis
Employed advanced sentiment analysis to identify emotional context in posts
Enhanced detection precision through innovative NLP and deep learning techniques

→ View Project

Cyberbullying Detection using Machine Learning

Classical ML approach for social media content moderation:

Applied Logistic Regression, SVM, and KNN algorithms for detecting cyberbullying
Conducted comprehensive comparative analysis of supervised and unsupervised approaches
Effectively identified and classified abusive language in social media content

→ View Project