Multi-Agent Framework to Augment Fraud Detection & Human Review

The rapid evolution of generative tools has enabled sophisticated ad fraud and phishing campaigns that outpace human review capacity. Traditional classifier-based detection systems struggle to explain, adapt, and scale against nuanced attacker tactics such as multi-redirect phishing chains, impersonation, and short-lived domains. This paper presents a multi-agent framework that uses LLM-based debate and evidence synthesis to augment fraud detection and human review in large-scale ad ecosystems. Our approach decomposes fraud investigation into specialized agents: Evidence Collector, Signal Enricher, Domain SME, Proponent, Opponent, Jury, and Summarizer, that collaborate to build structured, provenance-backed cases for each suspicious advertiser. The proponent and opponent agents debate enforcement decisions; a jury agent synthesizes confidence-weighted verdicts; and a summarizer translates these into concise, auditable rationales for human reviewers. This architecture standardizes inputs, mitigates bias, and ensures reviewers operate with a consistent and comprehensive evidence set. Early pilots in ad phishing detection demonstrate measurable gains: up to 40% faster review times, 20% higher inter-rater agreement, and significant recall lift on subtle fraud patterns that evade heuristic or rule-based systems. Moreover, debate-based deliberation enhances explainability and supports defensible appeal handling a critical need for regulated ad markets. By operationalizing multi-agent reasoning, the framework enables a scalable, transparent, and reviewer-aligned approach to combating fraud in the modern digital advertising ecosystem. It bridges the gap between AI explainability research and the operational realities of trust and safety review at global platform scale.

Some content is hidden, to be able to see it login here Login