Dr Yigitcan Kaya - UC Santa Barbara

AI is increasingly used to detect malware, prioritise alerts, and support cyber defence. Yet security environments are not static. Software ecosystems evolve, attackers adapt, organisations follow different risk policies, and the same behaviour can be harmless in one context but suspicious in another. As a result, AI security systems that perform well in benchmark settings can become fragile after deployment. 

My research studies this gap between controlled evaluation and operational reality. Across two recent works, I focus on two major sources of failure in AI-for-security systems. The first is policy discrepancy: the assumption that every security sample has one universal ground-truth label breaks down when labels depend on organisational context, such as whether a dual-use tool is acceptable or malicious. The second is heterogeneous concept drift: malware distributions change over time, but not all changes break models in the same way, so no single update strategy is reliable across all deployment conditions. 

To address these problems, I develop adaptive methods that make security models more aware of the conditions under which they operate. One line of work adapts deployed models using limited feedback from their target environment, aligning predictions with local security policies while preserving generalisation. Another line of work treats model updating as a deployment-time decision problem, selecting among adaptation strategies based on observable signs of distribution shift. 

Together, these works argue for a broader shift in AI-for-security research: from static models evaluated under simplified assumptions to adaptive systems that recognise ambiguity, diagnose change, and respond with the right level of supervision. The goal is not simply higher benchmark accuracy, but more reliable cyber defence under the messy, evolving conditions of the real world.