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

(Accepted) ACL Student Research Workshop (SRW) 2025

[DOI] [Slides]

When given multiple tasks at once, can LLMs effectively assess the safety of each task separately? In this work, I co-developed a novel blackbox jailbreak attack showing that LLMs tend to overlook harmful context while handling a series of prompts in a single query. While existing jailbreaks often require multiple rounds for success, our attack achieves a high success rate (~90% in JBB-Behaviors dataset) against recent LLMs with a single, simple query.
Our strategy was to embed a harmful prompt within a series of benign ones, causing the model’s attention to be diluted across prompts and reducing its focus on the harmful prompt. We showed three different example attack scenarios (Question bank, dialog completion , and level-wise game scenario description) where sequential prompt structure can be used to elicit harmful responses. This research was done in collaboration with Penn State University.

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

[Writing Sample]

As a follow-up to my previous collaboration, this is an ongoing project on a comprehensive survey examining the landscape of attacks and defenses targeting LLM agents and Vision-Language Models (VLMs). In this study, I aim to explore questions such as: What additional attack surfaces do multimodal models and AI agents introduce beyond those of unimodal LLMs? and To what extent and how can existing attacks and defenses for unimodal LLMs be adapted to multimodal models and agents?

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 ICASSP 2026

Undergraduate thesis project under Dr. Tanzima Hashem (BUET). Methane detection and reduction offers a rapid path to limiting global temperature rise. But what makes methane detection a challenging segmentation problem? First, publicly available datasets are scarce and highly imbalanced, with a very high negative-to-positive ratio due to the scarcity of recorded methane leakage events. Second, the complex morphology of gas plumes and satellite data noise make the segmentation task even more difficult.

Our contributions were mainly two: using U-Net with attention gates to dynamically focus on methane-specific regions, and demonstrating that spectral indices as an extra input channel can produce more accurate masks. We also address the high class imbalance in our dataset with focal loss. Our approach achieved a balanced performance by reducing both false positives and false negatives, which existing works could not achieve.

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. For binary classification tasks like this, clearly separating features helps models generalize better to hard-to-classify cases. Building on this, our methodology employs MaxViT, CoAtNet, and EVA‑02 as backbones, fine-tuned using supervised contrastive loss to enhance feature separation. The final prediction is obtained through majority voting on the classification heads of the individual backbones. Supervised contrastive loss produces well-separated embedding spaces, while ensembling captures different perspectives from multiple models.