Watch, Listen, Understand, Mislead: Tri-modal Adversarial Attacks on Short Videos for Content Appropriateness Evaluation
Supervised by: Dr. Swakkhar Shatabda.
Published at SVU Workshop ICCV 2025 [link]
In this paper, we introduce a comprehensive framework for evaluating the tri-modal safety of MLLMs. First, we present the Short-Video Multimodal Adversarial (SVMA) dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks. Second, we propose ChimeraBreak, a novel tri-modal attack strategy that simultaneously challenges visual, auditory, and semantic reasoning pathways. Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR). Our findings uncover distinct failure modes, showing model biases toward misclassifying benign or policy-violating content.

AVLoS: Audio-Visual Long Text Scene Summarization
Supervised by: Dr. Swakkhar Shatabda.
Thesis Work. In Progress.
We are experimenting on fine-grained long text scene summarization from videos using video and audio inputs. Most approaches use queries to guide text generation, we are experimenting to generate text without these queries.

MIMIC: Multimodal Islamophobic Meme Identification and Classification\
Supervised by: Nahid Hossain.
Accepted at MusIML Workshop - NeurIPS 2024. | [codebase] [arxiv]
Anti-Muslim hate speech has emerged within memes, characterized by context-dependent and rhetorical messages using text and images that seemingly mimic humor but convey Islamophobic sentiments. This work presents a novel dataset and proposes a classifier based on the Vision-and-Language Transformer (ViLT) specifically tailored to identify anti-Muslim hate within memes by integrating both visual and textual representations.

BhaShammo: IPA Transcription of Bengali Regional Dialect using Dialect Guided Tokens
Co-supervised by: Dr. Swakkhar Shatabda and Dr. Farig Sadeque.
In Progress. | [codebase] [arxiv]
We present an approach to transcribe regional Bengali text to IPA by introducing the Dialect Guided Tokens (DGT) technique on a new dataset spanning six districts of Bangladesh. We provide the model with information on the regional dialect of the input text before generating the IPA transcription. This is the first time this problem has been solved.

E-MedViTR: Enhanced Vision Transformers with Registers for Biomedical Image Classification

Supervised by: Dr. Dewan Md. Farid.
Published at ICEEICT 2024. | [paper]
We investigated the effectiveness of the ViT with registers (aka DINO v2) in classifying medical pathology images. Normally, it doesn't perform up to the mark as SOTA models do, but with an extension and data augmentations, it performs relatively close to SOTA models.

IPBlocks: A Blockchain Ecosystem for Secure IP Registration and Decentralized Marketplace
Supervised by: Dr. Salekul Islam 

Published at TENCON 2025
We introduce IPBlocks, a blockchain-based solution that streamlines IP applications, trading, and royalty transfers through a decentralized marketplace. We develop a set of algorithms that allow users to apply for, publish, auction, and transfer IPs.