Abstract

The integration of AI-driven forecasting systems into operational decision-making has introduced new vulnerabilities: adversarial agents can strategically manipulate the data these systems rely on, compromising their integrity and the decisions they inform. This paper develops an adversarial AI risk analysis framework for sequential Bayesian count data forecasting models under strategic manipulation, motivated by click fraud in online advertising — a domain where AI-powered bidding systems are prime targets for data poisoning. In adversarial settings, sequential Bayesian models are susceptible to attacks in which an adversary perturbs observations to manipulate filtering or forecasting distributions and disrupt downstream AI-driven decision-making. To better understand the underlying mechanics of such adversarial decisions, the paper introduces several formulations accounting for both attacker and defender perspectives across various attacker archetypes. Scenarios are considered in which perturbations propagate sequentially through both filtering and forecasting uncertainty mechanisms. The implications of these sequential adversarial effects on uncertainty quantification and resulting decisions are investigated through simulated and real click fraud examples involving single- and multi-step attack and defense strategies. The work contributes a statistically grounded lens for assessing and improving the robustness of AI forecasting systems against deliberate manipulation.