Discover AI Model Ontology
This is like a burglar figuring out exactly what kind of security system a house has by testing the doors and windows repeatedly to see what triggers an alarm. By watching how the AI responds to different inputs, an attacker learns what the system is 'looking for,' which helps them find a way to sneak past it.
The process of mapping an AI model's output space—such as its classification labels or detection categories—through iterative probing or analysis of exposed metadata. By enumerating the model's response patterns, an adversary gains intelligence on the system's functional scope, allowing them to tailor adversarial inputs to bypass specific security controls or exploit known model limitations.
The systematic reconstruction of an AI model's output ontology, defined as the set of discrete classes, labels, or latent features the model is capable of resolving. This discovery is achieved via black-box query enumeration, side-channel analysis, or the exfiltration of configuration artifacts. The resulting ontology map provides the adversary with the necessary semantic context to optimize adversarial perturbations, facilitating targeted evasion or poisoning attacks by aligning malicious inputs with the model's specific classification schema.