FraudBench: A Multimodal Benchmark for Detecting AI-Generated Fraudulent Refund Evidence
Researchers have introduced FraudBench, a new multimodal benchmark designed to detect AI-generated fraudulent refund evidence in e-commerce, food delivery, and travel services. As AI-generated images become increasingly realistic, they pose significant challenges for verifying visual claims of product damage or service defects. Existing detection benchmarks often overlook claim-conditioned fraud, focusing instead on standalone authenticity. FraudBench addresses this gap by curating real-world user-review evidence and synthesizing fake-damaged images using six state-of-the-art generation models. The study evaluates Multimodal Large Language Models (MLLMs), specialized detectors, and human participants. Results indicate that current MLLMs struggle significantly with detecting fake damage, often performing below a 50% baseline. While specialized detectors show better performance, they remain inconsistent across different generators and prone to false positives on genuine damaged items. This research highlights a critical disparity between generic AI image detection capabilities and the nuanced requirements for reliable, claim-specific fraud verification, underscoring the urgent need for improved tools to combat synthetic media misuse in commercial transactions.
Wire timeline
FraudBench: A Multimodal Benchmark for Detecting AI-Generated Fraudulent Refund Evidence
Researchers have introduced FraudBench, a new multimodal benchmark designed to detect AI-generated fraudulent refund evidence in e-commerce, food delivery, and travel services. As AI-generated images become increasingly realistic, they pose significant challenges for verifying visual claims of product damage or service defects. Existing detection benchmarks often overlook claim-conditioned fraud, focusing instead on standalone authenticity. FraudBench addresses this gap by curating real-world user-review evidence and synthesizing fake-damaged images using six state-of-the-art generation models. The study evaluates Multimodal Large Language Models (MLLMs), specialized detectors, and human participants. Results indicate that current MLLMs struggle significantly with detecting fake damage, often performing below a 50% baseline. While specialized detectors show better performance, they remain inconsistent across different generators and prone to false positives on genuine damaged items. This research highlights a critical disparity between generic AI image detection capabilities and the nuanced requirements for reliable, claim-specific fraud verification, underscoring the urgent need for improved tools to combat synthetic media misuse in commercial transactions.
cs.AI updates on arXiv.org