SECURITY AND AI LAB

Research

Research at Security and AI Lab (SAIL) covers cybersecurity, artificial intelligence, and computer science, with an emphasis on understanding emerging digital threats and developing reliable, practical approaches to security. Our work brings together methods from machine learning, data analysis, and intelligent systems to investigate topics such as malicious URL detection, AI-generated content, image protection, and malware analysis. The research areas below describe our main scientific interests and present related research work. The Research Projects section presents projects developed within these research areas. Together, these activities reflect how SAIL connects scientific investigation, academic collaboration, and the development of new research directions.

Malicious URL and Phishing Detection

This research area investigates methods for distinguishing legitimate web addresses from phishing, malware-related, and other malicious URLs. It examines URL representations, machine-learning approaches, feature efficiency, and the reliability of detection across different threat categories and datasets.

Published 2026 A feature-efficient hierarchical classification strategy for malicious URL detection Özdaş M. B., Osmanoğlu M. Journal of Computer Virology and Hacking Techniques This study approaches malicious URL classification as a sequence of binary decisions rather than a single multiclass prediction. It combines lexical and structural URL features with contextual representations, examines feature-selection strategies, and compares classification models on benchmark datasets to assess performance with more compact feature sets.View publication
Published 2026 A feature-enriched deep learning based ensemble framework for robust phishing URL detection Kaya D., Osmanoğlu M. PeerJ Computer Science This study investigates phishing URL detection using enriched structural and lexical features. It compares traditional machine-learning techniques with deep-learning and ensemble models, including combinations of handcrafted URL features and sequence-based representations, to assess classification performance and the contribution of feature design.View publication

LLM-Generated Text Detection

This area studies methods for distinguishing human-written text from text generated by large language models. Research examines signals in language-model predictions, approaches that do not require detector-specific training, and the consistency of detection across different datasets, writing domains, and generator models.

Preprint 2026 DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection Mehmet Batuhan Özdaş, Murat Osmanoğlu arXiv · 2026 DWT-Fusion studies how AI-generated text can be detected without training a separate supervised classifier. The framework analyzes token-level probability sequences using discrete wavelet transforms, combines signals across multiple scales, and evaluates detection behavior on different datasets and language models.View preprint
In Preparation

Supervised Wavelet-Feature-Based LLM-Generated Text Detection
Mehmet Batuhan Özdaş, Murat Osmanoğlu
This ongoing research explores supervised detection of LLM-generated text using multiresolution wavelet features extracted from token log-probability sequences. It studies feature representations derived from multiple wavelet families, compares compact and extended feature sets, and evaluates classifiers across datasets and proxy language models.

Image Scrambling and Embedded Security

This research examines image scrambling and lightweight protection techniques for resource-constrained platforms. The work includes reversible permutation frameworks, structural properties of image transformations, and experimental considerations for IoT and embedded devices.

Published2026Knit scrambling: A novel image scrambling framework and its demonstration in image encryptionKaya M. S., İnce K.Journal of Information Security and ApplicationsThis work introduces a reversible image scrambling framework based on permutation rules and evaluates its use within image encryption. It provides a research foundation for exploring resource-aware scrambling methods in embedded and IoT environments.View publication

Malware Detection and Digital Forensics

This area investigates static methods for recognizing malicious files without executing them. Research focuses on file-format characteristics, structural and statistical features, explainability, and detection robustness when tested on previously unseen malware families.

Published2026Static detection of malicious windows shortcut files without content-specific indicatorsMuhammed Saadetdin KayaForensic Science International: Digital InvestigationThe study investigates whether malicious Windows shortcut (LNK) files can be identified using static statistical features without relying on known malicious commands, URLs, or file paths. It reduces 73 input features to nine, examines their contributions, and evaluates detection on malware families excluded from training.View publication
RESEARCH INITIATIVES

Research Projects

SAIL research projects include approved studies, proposals under review, and initiatives in preparation. The project cards summarize their research aims, current status, and participating researchers.

Ankara University BAPApproved · Pending Start

Edge-Based Autonomous Intrusion Detection and Response for IoT Networks

This approved project will develop a hybrid edge-to-cloud architecture for intrusion detection and response in resource-constrained IoT and industrial IoT networks. Lightweight models analyze network traffic at the edge, explainable AI (XAI) and small language models (SLMs) prepare privacy-conscious security summaries, and multiple large language models support decision evaluation. The design includes policy-checked responses such as traffic limiting and device isolation.

Academic AdvisorAsst. Prof. Dr. Murat Osmanoğlu
Research TeamLect. Dr. Muhammed Saadetdin KayaLect. Merve Işık TarakcıLect. Mehmet Batuhan Özdaş
TÜBİTAK BİGG 1812Approved · Pending Start

Privacy-Preserving and Explainable AI for Architectural Project Review

This approved entrepreneurship initiative plans to develop a decision-support platform for preliminary checks of architectural drawings against applicable building and planning regulations. The proposed system analyzes CAD/BIM project documents, identifies architectural elements and spatial measurements, and reports potential non-compliance with references to relevant provisions. It is designed to complement, rather than replace, existing CAD/BIM tools.

Research TeamLect. Dr. Muhammed Saadetdin KayaNihal Çınar Kaya
TÜBİTAK 1001Under Review

Human Visual Attention for AI-Generated Video Detection

This proposal, currently under review, investigates whether visual attention patterns observed when people distinguish authentic from AI-generated videos can improve automated detection. It plans to estimate gaze from standard camera recordings, analyze attention differences across real and synthetic videos, and evaluate attention-informed classifiers using tests that hold out individual video generators to assess generalization.

Academic AdvisorProf. Dr. Mehmet Serdar Güzel
Research TeamLect. Dr. Muhammed Saadetdin KayaBurak KayaAssoc. Prof. Dr. Ahmet Çağdaş SeçkinAsst. Prof. Dr. Murat Osmanoğlu
TÜBİTAK 1002-AUnder Review

Human-in-the-Loop Visual CAPTCHA Development and Evaluation

The proposed study develops and evaluates AI-generated visual CAPTCHA challenges. Independent human reviewers assess visual quality and semantic suitability, while automated tests measure how difficult the challenges are for machine solvers. Evaluation feedback is then used to refine challenge generation, with security and human usability assessed together.

Academic AdvisorAsst. Prof. Dr. Murat Osmanoğlu
Research TeamLect. Mehmet Batuhan ÖzdaşLect. Dr. Muhammed Saadetdin KayaLect. Merve Işık Tarakcı