Publications

You can also find my articles on my Google Scholar profile

1. Plot-and-Ask: Multimodal Local LLMs in Smart Environments for Visual IoT Analytics

Authors: A. Varol, A. Shaikh, N.H. Motlagh, M. Leino, J. Virkki
Conference: 2026 21st Conference on Computer Science and Intelligence Systems (FedCSIS), ACSIS Vol. 47
Pages: 609-614
Year: 2026
Summary: This paper presents a local multimodal analytics architecture for smart indoor environments that combines edge-resident multimodal large language models (MLLMs) with a knowledge graph (KG) policy layer. The proposed plot-and-ask workflow turns longer sensing periods into time-series plots with KG-derived comfort bands and passes them to a local vision-enabled model for interpretation, giving small models a compact representation of the data within their limited context windows. In a two-week indoor air quality deployment, all four evaluated local models achieved 100% accuracy for natural-language-to-SQL synthesis and plot generation, and Ministral 3 3B reached 100% plot-interpretation accuracy with 5.84 s end-to-end latency, with all processing kept on local hardware.
Link: View Publication
Code: GitHub

2. Multi-Agent Large Language Models for Distributed Internet of Things Analytics in Smart Environments

Authors: A. Varol, A. Shaikh, N.H. Motlagh, M. Leino, J. Virkki
Conference: 2026 11th International Conference on Smart and Sustainable Technologies (SpliTech)
Year: 2026
Summary: This paper presents a hierarchical multi-agent architecture in which a supervisor agent on an edge device delegates IoT analytics tasks to location-specific worker agents running on Raspberry Pi devices, with Retrieval-Augmented Generation (RAG) grounding the outputs in sensor measurements. Deployed with BME680 sensors in an office, a kitchen, and a hallway at Tampere University, the system achieves 100% accuracy across 135 queries under normal operation, retains 88.9% with one agent offline, and drops to 62.2% when a single compromised agent returns fabricated values, highlighting trust-aware aggregation as a key design requirement for multi-agent smart environments.
Link: View Publication

3. Evil-AI Benchmark: An Evaluation Framework for Adversarial Security Assessment of LLM Agents in Smart Environments

Authors: A. Varol, A. Shaikh, H. Jalo, G. Muchenje, J.S. de Cerqueira, K.-K. Kemell, N.H. Motlagh, M. Leino, H. Pirkkalainen, P. Abrahamsson, J. Virkki
Journal: IEEE Access, Vol. 14
Pages: 133098-133116
Year: 2026
Summary: This paper introduces Evil-AI Benchmark, an open-source framework for assessing adversarial threats to LLM agents in smart environments. It evaluates five threat vectors—prompt injection, persuasion, attacker-in-the-middle simulation, data leakage, and unsafe actions—alongside device-control capability and over-refusal, including physical verification of actuator commands. Eight LLMs are tested using 350 scenarios, including 250 adversarial tests. Observed attack success rates range from 0.8% to 54.0%, with persuasion and prompt injection presenting persistent weaknesses. The benchmark provides reproducible safety metrics for cyber-physical LLM applications.
Link: View Publication

4. Measuring the Atmosphere: Environmental Sensing for Occupant Flows in Commercial Spaces

Authors: N.H. Motlagh, A. Varol, M.A. Zaidan, P.L. Fung, S. Varjonen, P. Nurmi, S. Tarkoma
Journal: IEEE Internet of Things Journal
Year: 2026
Summary: This paper investigates privacy-preserving visitor-flow estimation using environmental sensors deployed across a commercial building. Fourteen multivariable sensor nodes collect data in a shopping mall for up to 79 days, with ground-truth visitor counts available at one location. A compact multimodal attention-based model fuses environmental signals to classify low, medium, and high occupancy, achieving 76.5% balanced accuracy under a temporally blocked data split. The model outperforms seven baselines while using only 5,813 parameters (approximately 23 kB), demonstrating the potential for resource-constrained edge inference.
Link: View Publication

5. A Structured Prompt-Based Approach to Evaluating Exploitative Misuse in Generative AI

Authors: I. Savolainen, J. Vestelin, T. Ihalainen, A. Varol, J. Virkki
Conference: 2026 11th International Conference on Smart and Sustainable Technologies (SpliTech)
Year: 2026
Summary: This paper examines a structured prompt-based approach to evaluating exploitative misuse risks in generative AI. It focuses on using systematically constructed prompts to investigate potentially harmful uses of generative models and inform AI safety assessment.
Link: View Publication

Authors: T. Vuohijoki, J. Huhtasalo, M. Raukola-Lindblom, R. Deloso, M.M. Sakif, O. Kivin, T. Rautio, A. Tavakoli, C. Gonzalez Visiedo, H. Mattila, S.M. Rahman, P. Sipola, T. Ihalainen, A. Varol, J. Virkki, S. Merilampi
Conference: 2026 11th International Conference on Smart and Sustainable Technologies (SpliTech)
Year: 2026
Summary: This paper addresses user-centred concept development for cloth-based interactive activators intended to support social inclusion. It explores the potential of textile-based interaction concepts to encourage participation and help make everyday environments and activities more accessible.
Link: View Publication

7. Creation of AI-driven Smart Spaces for Enhanced Indoor Environments – A Survey

Authors: A. Varol, N.H. Motlagh, M. Leino, S. Tarkoma, J. Virkki
Journal: Internet of Things, Vol. 36, Art. 101876
Year: 2026
Summary: This survey introduces foundational components required to build AI-driven smart environments, emphasizing the integration of emerging AI methodologies to enhance functionalities such as personalized comfort settings and interactive living spaces. It provides a comprehensive overview of sensor technologies, data communication protocols, and the role of AI in creating intelligent, adaptive indoor environments.
Link: View Publication

8. Performance of Large Language Models Across Edge and Cloud Platforms in Smart Spaces

Authors: A. Varol, N.H. Motlagh, M. Leino, J. Virkki
Conference: 2025 10th International Conference on Smart and Sustainable Technologies (SpliTech)
Year: 2025
Summary: This paper presents a hybrid edge–cloud framework for deploying large language models (LLMs) in smart spaces, integrating open-source models on edge devices (laptops, Raspberry Pis) with a commercial cloud platform. It evaluates performance in terms of latency and accuracy across descriptive, diagnostic, predictive, and prescriptive analytics using both historical and real-time environmental data. Findings highlight the complementary strengths of cloud-based and edge-based deployments, with a task-aware split improving performance, privacy, and resilience in AI-driven smart environments.
Link: View Publication

9. From Prompts to Motors: Man-in-the-Middle Attacks on LLM-Enabled Vacuum Robots

Authors: A. Shaikh, A. Varol, J. Virkki
Journal: IEEE Access, Vol. 13
Pages: 137505-137513
Year: 2025
Summary: This paper presents a three-phase testbed and gray-box threat model for man-in-the-middle (MITM) attacks on an LLM-enabled vacuum robot. The presented system converts natural-language prompts into motor commands. The study evaluates four attacks—two indirect prompt-injection and two output-manipulation scenarios—across GPT-4, GPT-4o mini, and GPT-3.5 Turbo in simulated, tabletop, and physical trials, showing that a remote adversary who intercepts JSON messages can bypass collision avoidance, override motor control, and deliver deceptive feedback to users.
Link: View Publication

10. A single-IC realizable, electronically tunable, OTA-based full-wave rectifier with simultaneous positive and negative outputs

Authors: A. Varol, F. Yucel, E. Yuce, A. Cakir
Journal: AEU-International Journal of Electronics and Communications, Vol. 183, Art. 155374
Year: 2024
Summary: This paper presents a novel operational transconductance amplifier (OTA)-based full-wave rectifier circuit that is electronically tunable and capable of providing simultaneous positive and negative outputs. The design is suitable for integration into a single integrated circuit (IC), offering advantages in analog signal processing applications.
Link: View Publication

11. LeNet ve ResNet Derin Öğrenme Modelleri ile Asma Yapraklarının Sınıflandırması

Authors: K. Kayaalp, A. Varol
Journal: Veri Bilimi, Vol. 7, No. 1
Pages: 16-25
Year: 2024
Summary: This study explores the classification of grapevine leaves using deep learning models, specifically LeNet and ResNet architectures. The research demonstrates the effectiveness of these models in accurately classifying different types of grapevine leaves, contributing to advancements in agricultural applications of deep learning.
Link: View Publication

12. A new electronically tunable transimpedance-mode OTA-based first-order universal filter and its quadrature oscillator application

Authors: A. Varol, F. Yucel, A. Cakir
Journal: Journal of Circuits, Systems and Computers, Vol. 32, No. 11
Year: 2023
Summary: This paper introduces a novel transimpedance-mode first-order universal filter using operational transconductance amplifiers (OTAs). The proposed circuit is electronically tunable and can function as a quadrature oscillator, offering versatility in analog signal processing applications.
Link: View Publication

13. A New Mobile Application for Physical Measurement in a Cellular Network

Authors: A. Varol, B.K. Çetin
Journal: Journal of Scientific Reports-A
Pages: 178-200
Year: 2020
Summary: This research presents a mobile application designed for physical measurements within cellular networks. The application aims to provide accurate and efficient measurement capabilities, enhancing the analysis and optimization of cellular network performance.
Link: View Publication

Preprints

Enabling Cloud-Level Accuracy in Edge AI through IoT Data Preprocessing

Authors: A. Varol, K. Kołodziej, Ł. Sobczak, M. Romaszewski, P. Głomb, N.H. Motlagh, M. Leino, J. Virkki
Preprint: arXiv:2606.22496
Year: 2026
Summary: This paper examines whether prompt preprocessing can help small local LLMs interpret IoT air-quality and thermal-comfort data nearly as well as cloud models. Raw sensor readings are converted into progressively richer text, from raw values to threshold-aware descriptions and summary flags, and five local and five cloud LLMs are evaluated on indoor and outdoor datasets with and without chain-of-thought prompting. Prompt enrichment substantially improves local-model accuracy, showing that lightweight preprocessing can narrow the gap between local and cloud performance.
Link: View Preprint

Aygün Varol