Blog
Thoughts on data science, ML engineering, and AI research
Adversarial Attacks on Fraud Detection: What My Thesis Found
My MSc thesis set out to answer a dangerous question: what happens to XGBoost fraud detection models and their SHAP explanations when a sophisticated adversary deliberately crafts transactions to evade detection? The answer was worse than expected.
Why SHAP Explanations Break Under Adversarial Pressure
The EU AI Act classifies fraud detection systems as high-risk AI. Article 13 requires meaningful explanations. But what if the explanations themselves can be manipulated by the very adversary you're trying to detect? This is the regulatory gap my research addresses.
SWIFT & SEPA Payments: How AI Can Catch What Rules Miss
Rule-based systems flag what they've seen before. Machine learning models catch what rules miss. But neither alone is enough for high-value cross-border payments where milliseconds and millions are both at stake. Here's how I built a hybrid detection pipeline.