Data representation is the structured form used to encode real-world events or observations as information a computer system can store and process. It may include fields, labels, categories, timestamps, logs, or API records. The representation determines which details are captured and how systems can interpret them.
In AI security, representation matters because a model can only assess information that reaches it in a usable form. A missing field, narrow category, or lossy transformation can hide relevant evidence without causing an obvious technical error. Attackers may also exploit weak representations, for example by crafting inputs that are misclassified or evade detection. Defenses include reviewing schemas and data transformations, documenting limitations and provenance, and testing whether important evidence can enter the workflow and be surfaced for human review.