Research.
My PhD measures the gap between how sure an AI sounds and what the evidence behind it actually supports.
What is exploitable. What is malicious. Who is responsible.
Same confidence, different evidence.
AI systems increasingly make those calls in security. They report their conclusions with the same confidence whether the evidence behind them is conclusive or thin. My thesis measures that gap.
I treat it as a calibration problem, not a capability problem. So the contribution isn’t a better attacker or a better detector. It’s ways to measure the gap, and evidence about it.
The official title
“AI solutions for cybersecurity and methods to secure AI-based deployments”
PhD, Transilvania University of BrașovThree questions.
Three security settings where an AI verdict carries weight. Together they move from reproducibility, to calibration, to auditability.
- RQ1Discovery
Is an AI-found vulnerability real?
Run the same task again. Does the finding still hold up, or was it a one-off?
Tests reproducibility - RQ2Trustworthiness
How far can the answer be believed?
Telling apart answers that are wrong but confident from answers that have been manipulated.
Tests calibration - RQ3Reconstruction
After a failure, can we recover what happened?
Recovering the truth, and what the system knew at the time it made the call.
Tests auditability