Trustworthy AI

Computer Vision

Deep Learning

SequentialBreak: Large Language Models Can be Fooled by Embedding Jailbreak Prompts into Sequential Prompt Chains (2024-2025)

Bijoy Ahmed Saiem*, MD Sadik Hossain Shanto*, Rakib Ahsan*, Md Rafi ur Rashid

[DOI]

With collaboration with Md Rafi Ur Rashid (Penn State University), we developed a novel automated jailbreak attack for LLMs. Our methodology exploited the attention dilution phenomenon by embedding the harmful prompt within a series of benign prompts in a single query. We showed three different example attack scenarios where sequential prompt structure can be used to elicit harmful responses. While existing jailbreak methods often require multiple rounds for success, our one-shot attack achieved a substantially high attack success rate against recent LLM models like Llama-3 and GPT-4o.

A Survey of Attacks and Defenses in LLM Agents and Vision‑Language Models (2025-Current)

Bijoy Ahmed Saiem*, MD Sadik Hossain Shanto*, Raihan Sobhan*, Rakib Ahsan*, Md Rafi ur Rashid

(Accepted) ACL Student Research Workshop (SRW) 2025

Ongoing project on a comprehensive survey in collaboration with Md Rafi Ur Rashid (Penn State University), examining the landscape of attacks and defenses targeting large language model (LLM) agents and vision-language models. The survey covers a broad spectrum of security threats, including adversarial attacks, jailbreaks, data poisoning, backdoors, privacy breaches, and model extraction attacks.

AttMetNet: Attention-Enhanced Deep Neural Network for Methane Plume Detection in Sentinel-2 Satellite Imagery (2024-2025)

Undergraduate Thesis

Rakib Ahsan*, MD Sadik Hossain Shanto*, Md Sultanul Arifin, Tanzima Hashem

Under preparation for "NeurIPS 2025 Workshop: Tackling Climate Change with Machine Learning"

Undergraduate thesis project under Dr. Tanzima Hashem (BUET). Methane is the second-largest contributor to greenhouse gas emissions, and its reduction offers a rapid path to limiting global temperature rise. Our aim in this work is to develop a lightweight framework that outputs accurate segmentation masks of methane plumes from multispectral satellite imagery. For segmentation model, we enhanced U-net with attention gates for improved focus in methane-specific region. We also showed that introducing spectral indices as an extra input channel significantly improves segmentation boundaries. Experiments show that our method outperforms the state-of-the-art detection methods in real-world conditions.

DFCon: Attention-Driven Supervised Contrastive Learning for Robust Deepfake Detection (2024)

MD Sadik Hossain Shanto*, Mahir Labib Dihan*, Souvik Ghosh*, Riad Ahmed Anonto*, Hafijul Hoque Chowdhury*, Abir Muhtasim*, Rakib Ahsan*, MD Tanvir Hassan*, MD Roqunuzzaman Sojib*, Sheikh Azizul Hakim, M. Saifur Rahman

[arxiv]

This work originated from our submission to the IEEE SPS Signal Processing Cup 2025, where our team (Straw Hats) was the 1st runner-up. The task involved classifying human faces as deepfake or real. Our approach utilized MaxViT, CoAtNet, and EVA‑02 as backbone models, fine-tuned with supervised contrastive loss to enhance feature separation. Final predictions were obtained through a majority voting ensemble across the backbone classifiers.