My research and Industrial projects covers the topics of -

• Machine Learning
• Deep Learning
• Natural Language Processing
• Large Language Models
• Computer Vision
• Cybersecurity

1. Thesis

   Investigation and Transforming Resource-Scarce Language Question Generation: A Study on Text-to-Text Transformer Models for Bengali

   Fahad Siddique Faisal

   Undergraduate Thesis, Chittagong University of Engineering and Technology, 2024

   Developed and improved an automatic FAQ generation system using text-to-text transformer models to enhance Question Generation (QG) in multilingual contexts.

1. IEEE

   Automated Chest X-Ray Report Generation Using Vision-Language Models: A LLaMa 3.2 Based Approach

   M. A. Shovon, S. B. Emon, J. A. Sameha, F. S. Faisal, M. R. Sayem

   IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (IEEE QPAIN), 2026

   Accepted! Tell me about this paper

   Automated chest X-ray report generation significantly contributes to enhanced diagnostic efficiency and reduced workloads for radiologists through intelligent interpretation of medical imaging. This paper presents a framework for generating clinically coherent radiology reports from X-ray images.

   Combined visual understanding and linguistic reasoning using a LLaMA 3.2 Vision-Instruct model trained with EfficientNetB7 classifiers and parameter-efficient LoRA. Generated reports demonstrated clinical medical precision and diagnostic fluency with a structure and flow comparable to expert-authored text, highlighting the effectiveness of combining image classification with multimodal language modelling for reliable radiology report generation.

   
2. IEEE

   A Federated and Explainable Machine Learning Framework for Robust Intrusion Detection and Network Security Enhancement in Industrial Internet of Things (IIoT) Environments

   I. Das, G. Ali, A. Hossain, F. S. Faisal

   International Conference on Electrical, Computer & Telecommunication Engineering (IEEE ICECTE), 2026

   Accepted! Tell me about this paper

   Addressing the security of an organization from internal zero-day threats is highly problematic for classical security approaches, as such vulnerabilities are unknown and evade conventional detection methods. This paper aims to create a robust, portable ML-based intrusion detection system automated for zero-day attack detection under latency constraints.

   Developed a comprehensive preprocessing framework addressing missing values, feature selection, and dataset balancing with SMOTE. Evaluated Logistic Regression, Random Forest, XGBoost, LightGBM, Neural Networks, and CNNs. LightGBM and deep learning models demonstrated the best performance, justifying the need for advanced zero-day intrusion prevention.

   
3. IEEE

   Optimizing Resource-Scarce Language Question Generation on Bengali: A Comparative Study of Transformer-Based Models

   F. S. Faisal, S. N. Islam, M. F. Hossain, M. A. Hossain

   27th International Conference on Computer and Information Technology (IEEE ICCIT), 2024

   Paper Tell me about this paper

   This paper presents an automatic FAQ generation system for Bengali, a resource-scarce language, addressing limitations in existing Question Generation systems such as poor machine-translated datasets and inadequate evaluation metrics.

   Employed a series of text-to-text encoder-only and encoder-decoder transformer-based models (IndicBart, SahajBart, Biswabangla, BanglaBERT, BanglaT5, and mT5-base) to design the system. The proposed bn-TOK + mT5-base + Beam Search model scores Rouge 37.225 and SacreBLEU 14.202 with METEOR 0.216, demonstrating superior performance in balancing lexical accuracy, structural coherence, and semantic meaningfulness.

   
4. IEEE

   Enhancing Bilingual Sarcasm Detection with an Advanced Ensemble Model: A Dual-Language Deep Learning Framework

   S. N. Islam, F. S. Faisal, M. F. Hossain, M. A. Hossain

   27th International Conference on Computer and Information Technology (IEEE ICCIT), 2024

   Paper Tell me about this paper

   Sarcasm detection is crucial for interpreting web portal comments, customer feedback, and social media interactions, as misinterpretation can distort sentiment analysis. This study addresses the challenge of detecting subtle and context-dependent sarcastic expressions in both Bengali and English.

   Proposed an ensemble model integrating both BiLSTM and GRU with word embeddings. The model demonstrated superior performance compared to other deep learning models, achieving approximately 97% accuracy on the news headline dataset and 89% on the BanglaSARC dataset.

   
5. IEEE

   Design and Comparative Analysis of 1-BIT ALU and Full Adder using MGDI Technique for High-Performance PLC Applications

   M. S. H. Sarker, F. S. Faisal, A. Istiaque, T. Taher

   27th International Conference on Computer and Information Technology (IEEE ICCIT), 2024

   Paper Tell me about this paper

   With the growing demand for low-power, smaller chips, arithmetic operations performed by Arithmetic Logic Units and Full adders remain crucial in GPUs, CPUs, and microcontrollers. This paper addresses the need to reduce chip area and power dissipation in conventional programmable logic controller circuits.

   Designed a 1-BIT ALU, 10-T Full-adder, and NAND output selected with a multiplexer circuit using the Modified Gate Diffusion Input (MGDI) technique in a 90nm transistor size. The conventional PLC unit consists of at least 128 transistors, whereas the proposed design has only 54 transistors, significantly decreasing chip area and power dissipation.

   

1. Research

   CLHF-Med-AI: Medical Conversational AI Model Development with Continuous Learning (CLHF)

   Fahad Siddique Faisal (Author)

   Ongoing Research, Jul 2025 - Present

   High-performance medical assistant based on the Qwen2-0.5B-Instruct architecture. It is optimized via 4-bit QLoRA and implements a sophisticated continuous learning pipeline that refines model behavior through real-time human preference data.