Model drift

Model drift refers to the gradual degradation of an artificial intelligence system's performance over time, when the reality it encounters in production diverges from the data on which it was trained. A model learns the regularities of a snapshot of the world, but the world changes, behaviors evolve, distributions shift, and what was true at training gradually ceases to be so. A distinction is often drawn between data drift, when the characteristics of the inputs change, and concept drift, when the relationship between the inputs and the target to be predicted changes. The danger of drift is its stealth, the model continuing to produce normal-looking answers while becoming gradually less reliable, with no obvious alert. It therefore requires continuous monitoring of performance and periodic retraining. In cyber, this issue is particularly acute, because the threat evolves rapidly, attackers constantly adapting their techniques, which causes a rapid drift of detection models, which can become blind to attacks of a new kind. Drift thus constitutes an operational and liability risk, since an unmonitored system can cause harm long after it has ceased to be reliable.

A fraud-detection model trained on behaviors from two years ago becomes gradually less effective as fraudsters change their methods. For lack of monitoring, its drift goes unnoticed until a wave of frauds slips past it.

dérive du modèle, model drift, data drift, dérive des données, dégradation du modèle