Towards a Unified Abstract Architecture to Coherently and Generically Describe Security Goals and Risks of AI Systems.

We propose a unified abstract architecture for describing IT security goals and risks within AI systems. The architecture facilitates effective interdisciplinary communication among AI developers, data scientists, and security professionals. The architecture is abstract enough to cover a wide range of AI methods (not limited to machine learning) while it can still be used to sufficiently describe and map existing AI-specific attacks. It emphasises the importance of identifying at-risk processes or at-risk data within the AI system for a targeted increase of the overall system’s security. This systematic approach could help to optimise resource allocation while achieving desired protection targets for AI systems.