Trustworthy AI for Mission-Critical Systems
Machine learning research scientist and senior software engineer: trustworthy AI for mission-critical systems. U.S. Ph.D., three first-author IEEE Access papers.
Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
Field Notes
Field notes on ML security, proof-of-learning, distributed systems, and high-reliability software for researchers, engineers, and technical leaders.
Posts by Category
Posts organised by category: machine learning, security, engineering, and more.
Posts by Collection
Browse all publications, talks, and portfolio items on ozgurural.github.io.
Essays
Slow, indirect writing about machine learning, security, and the systems we are quietly building.
Who improved the AI?
Five organisations train one AI model. Who actually helped? A short visual story about evaluation, contribution rewards, privacy and evidence of training, grounded in Ural and Yoshigoe’s IEEE Access survey.
Hearing an attack in the noise
From Turkish posts and news to alerts for analyst review: a film on keyword selection, language normalisation and unusual activity around named targets (Ural & Acarturk, ICISSP 2021).
AI meets the deadline
An AI can plan a robot’s next move. What if the answer arrives after it was needed? A narrated film on keeping an AI planner out of a real-time controller’s waiting path. An open research direction of the author, not yet published.
A walk downhill
How does a machine actually learn? A narrated film on gradient descent: the step, why its size matters, momentum in narrow valleys, and saddle points.
Inside a Level D flight simulator
A narrated film by Dr. Ozgur Ural: how a Level D full-flight simulator fools a pilot’s senses, how it is proven against the real aircraft, why it can never be late, and what that means for AI in the loop.
Could this AI be a copy?
Could you recognise an AI model after its weights change? A short visual story about hidden patterns, chance matches and why model-security tests need calibration. An explicitly labelled statistical teaching model.
A backup that shares your mistake
Three computers agree. Could all three still be wrong? A narrated film on majority voting, common-cause failure and the loss of Ariane 5, whose two identical navigation computers failed the same way.
SecurePoL, animated
Someone copies your AI model. How do you prove it was yours? A two-minute narrated film on SecurePoL (Ural & Yoshigoe, IEEE Access 2025): keep the trail that training leaves, and plant a secret in the model that only its owner can ask for.
Signing an AI model
Your AI model has been stolen and runs inside someone else’s product. How do you show it is yours? A narrated film on model watermarking: three places to hide a signature in a model, and why it should be tied to the training record (Ural & Yoshigoe, IEEE Access 2025).
Who Improved the AI? Collaborative Learning
What a ledger buys machine learning: consensus that trains instead of hashing, contracts that pay for the improvement you caused, and the measured limits.
From Local Reports to Security Alerts
Turn Turkish posts and news into an investigation queue: learn relevant terms, normalise the language and flag unusual attention around a target.
AI Meets the Deadline
Explore how AI planning meets real-time execution: an animated architecture experiment with latency, stale proposals, fallback and resource contention.
Gradient Pinball: How Machines Learn
A cinematic explainer on gradient descent, momentum, and why high-dimensional loss landscapes are ruled by saddle points, not local minima.
Universal Jira: Can a Market Coordinate AI Agents?
An open research note: when AI agents both propose work and are paid for it, how could a prediction market tell useful work from reward hunting?
Inside a Level D Flight Simulator
How a Level D flight simulator fools a pilot’s senses, is proven against the real aircraft and is never late, and what that means for AI in the loop.
ML Oracles: Can a Contract Trust an AI Answer?
Can a contract trust an AI answer it did not compute? An animated comparison of zero-knowledge proofs, optimistic challenges, and their limits.
SecurePoL: Proving How an AI Model Was Trained
How Proof-of-Learning checks training records, where spoofing attacks exploit it, and how SecurePoL combines trajectory and watermark verification.
Block Race: Nakamoto Consensus
Why ‘6 confirmations’ is a probability, not a promise. A walk through the double-spend math of Bitcoin’s whitepaper §11.
Redundancy Reactor: When Three Computers Fail
Majority voting buys superlinear safety, until a shared cause installs a floor that more copies cannot remove. An explainer ending in the Ariane 5 loss.
Signing an AI Model: Three Kinds of Model Watermark
A narrated film on model watermarking: three places to hide a signature in a model (its answers, its weights, an auxiliary head), and why the mark should be tied to the training record.
Could This AI Be a Copy? Model Watermarks
How a watermark too faint to see in any single weight can become detectable across thousands of them under an explicit statistical model. A cinematic Z-test explainer.
Research Lab
Short films on who trained an AI model, learning across organisations and AI in real-time systems, each labelled as research, design study or teaching film.
Personal
Personal notes by Dr. Ozgur Ural: the Ph.D. years in Florida, living in the US and the Netherlands, and sailing with the METU team.
Projects
Research code, industry systems and applied AI work by Dr. Ozgur Ural: model provenance, real-time flight simulation, UAV mission control, enterprise AI architecture.
Publications
Peer-reviewed publications by Dr. Ozgur Ural on proof-of-learning, model watermarking, adversarial robustness, and blockchain-enhanced machine learning.
Advisory & Service
Technical advisory, ML security and model provenance review, conference speaking, peer review, and program-committee service.
Sitemap
A full list of all posts, pages, and publications on ozgurural.github.io.
Privacy and Site Use
How this portfolio handles hosting, external services, and contact messages.
Posts
Flight Simulation and the Evaluation of Learning Systems
General engineering considerations for evaluating learning systems in simulation, with a proposed research direction and its evidence limits.
METU Sailing Team: Racing in Urla, Bodrum, and Marmaris
Reflections on METU Sailing Club: racing on the Aegean coast in Urla, Bodrum, and Marmaris, and what competitive sailing taught me about teamwork.
Y Combinator and DNA: Fixed Points in Code and Biology
It started while I was nursing a Saturday-morning coffee and catching up on YouTube. Around the 11:20 mark of this talk, the speaker described the choreograp…
Proof of Learning: Building Trust in Future Machine Learning
Proof of Learning (PoL) verifies that a model was genuinely trained on claimed data by providing verifiable evidence of the training process. I first felt th…
Doctoral Research in Proof-of-Learning Security
My 2025 doctoral research at Embry-Riddle on spoofing attacks, model watermarking, and the limits of experimental evidence.
Model Watermarking and the Future of Trustworthy AI
Model watermarking embeds identifiable patterns into a model’s parameters or outputs so that ownership can be demonstrated without access to the original tra…
Blockchain Oracles and Smart Contracts: The World’s Agile Board
Traditional agile boards depend on centralized tools and manual updates. As projects span multiple organizations and trust boundaries, these boards struggle …
Blockchain-Enhanced Machine Learning: Linking Trust, Data, and Incentives
How blockchain’s immutable ledgers and decentralized governance address auditability, provenance, and incentive-alignment challenges in machine-learning pipelines.
Mission-Control Engineering and Autonomous Systems
Lessons and research questions from mission-control software work, with clear boundaries between engineering experience and future autonomy research.
USA vs Netherlands: An Expat’s Perspective
Living as an expat has given me the chance to compare daily life in the United States and the Netherlands in a way that guidebooks never could. My notebook i…
Adversarial Examples for Proof of Learning
Code for paper ‘Adversary examples’ for Proof of Learning
SecurePoL with Watermarking
Official code & experiments for my PhD dissertation: ‘Enhancing Security of Proof-of-Learning Against Spoofing Attacks Using Advanced Model Watermarking.’
Image Measurement with PyQt5: A Cardiology Aid
A clinician-calibrated EKG image measurement tool, and how its verified measurements can support human-in-the-loop clinical computer vision.
Presenting at ICISSP 2021: rare-event detection in a low-resource language
The talk behind the ICISSP 2021 paper: what an agglutinative language does to keyword detection, and what the sample run measured.
MS Thesis
My Master of Science Thesis Project related researches, projects and documents.
Secure Proxy on Cloud Paper Demonstration
Video demonstration of the Secure Proxy on Cloud paper, covering architecture motivation, implementation details, and development roadmap.
publications
Clover: Autonomous Cargo and Mail Delivery
Published in Turkish Autonomous Robots Conference (TORK 2014), 2014
Undergraduate capstone project demonstrating an autonomous cargo and mail delivery robot, presented at the Turkish Autonomous Robots Conference.
Citation
Yamaç, A. Ö., Doçi, J., Ural, O., Özarslan, Ö. F. and Temizer, S. (2014). Clover: Autonomous Cargo and Mail Delivery. Turkish Autonomous Robots Conference (TORK 2014), Ankara, Turkey.
Secure Proxy on Cloud
Published in Technical Report (ResearchGate), 2016
Technical report on a secure proxy architecture for cloud-based services.
Citation
Ural, O. and Erdur, E. (2016). "Secure Proxy on Cloud." Technical report. DOI: 10.13140/RG.2.2.24058.08649.
Automatic Detection of Cyber Security Events from Turkish Twitter Stream and Turkish Newspaper Data
Published in Master's Thesis, 2019
Master’s thesis introducing NLP methods for detecting cyber-security events in Turkish social media and news streams, advised by Prof. Cengiz Acartürk and the basis of the later ICISSP 2021 paper.
Citation
Ural, O. (2019). Automatic Detection of Cyber Security Events from Turkish Twitter Stream and Turkish Newspaper Data. Master's Thesis, Middle East Technical University, Ankara, Turkey.
Automatic Detection of Cyber Security Events from Turkish Twitter Stream and Newspaper Data
Published in Proceedings of the 7th International Conference on Information Systems Security and Privacy (ICISSP), 2021
A workflow for turning Turkish posts and news into alerts for analyst review: select relevant terms, normalise the language and track unusual activity around named targets.
Citation
Ural, O. and Acartürk, C. (2021). "Automatic Detection of Cyber Security Events from Turkish Twitter Stream and Newspaper Data." In Proceedings of the 7th International Conference on Information Systems Security and Privacy (ICISSP), pp. 66-76. DOI: 10.5220/0010201600660076.
Survey on Blockchain-Enhanced Machine Learning
Published in IEEE Access, 2023
Comprehensive survey of blockchain-enhanced machine learning: consensus-driven data provenance, federated learning on-chain, and incentive mechanisms across 120+ papers.
Citation
Ural, O. and Yoshigoe, K. (2023). Survey on Blockchain-Enhanced Machine Learning. IEEE Access, vol. 11, pp. 145331-145362. DOI: 10.1109/ACCESS.2023.3344669.
Enhancing Security of Proof-of-Learning Against Spoofing Attacks Using Feature-Based Model Watermarking
Published in IEEE Access, 2024
Feature-based model watermarking couples an ownership check to a Proof-of-Learning training record to study defences against spoofing.
Citation
Ural, O. and Yoshigoe, K. (2024). Enhancing Security of Proof-of-Learning Against Spoofing Attacks Using Feature-Based Model Watermarking. IEEE Access, vol. 12, pp. 169567-169591. DOI: 10.1109/ACCESS.2024.3489776.
Enhancing Proof-of-Learning Security Against Spoofing Attacks Using Model Watermarking
Published in Doctoral Dissertation, 2025
Doctoral research developing SecurePoL, a dual-layer framework coupling Proof-of-Learning trajectory logs with three orthogonal watermarking strategies.
Citation
Ural, O. (2025). Enhancing Proof-of-Learning Security Against Spoofing Attacks Using Model Watermarking. Doctoral dissertation, Embry-Riddle Aeronautical University.
SecurePoL: Integration of Watermarking With Proof-of-Learning to Enhance Security Against Spoofing Attacks
Published in IEEE Access, 2025
Dual-layer framework coupling immutable Proof-of-Learning logs with three watermarking strategies, so verification succeeds only when both the training trajectory and the watermark are consistent.
Citation
Ural, O. and Yoshigoe, K. (2025). SecurePoL: Integration of Watermarking With Proof-of-Learning to Enhance Security Against Spoofing Attacks. IEEE Access, vol. 13, pp. 213067-213091. DOI: 10.1109/ACCESS.2025.3642198.