Technical Summary: Generative AI-Enabled Refund Fraud in Chinese E-Commerce

1. Problem Statement

The paper addresses a critical security failure in Chinese e-commerce dispute resolution systems. Traditionally, these systems operate on the security assumption that digital evidence (e.g., photos of defective products) provided by buyers truthfully reflects physical reality. This assumption underpins automated "Refund Only" policies, where refunds are issued without requiring the return of physical items to optimize efficiency and reduce reverse logistics costs.

The rapid evolution of Generative AI (GenAI) has invalidated this threat model. Attackers can now utilize widely available GenAI tools to fabricate hyper-realistic, physically plausible evidence of non-existent defects (e.g., mold on fresh food, cracks on screens) at negligible cost. This capability transforms refund fraud from isolated, opportunistic acts into scalable, systematic operations. The core problem is the asymmetry of cost: the cost for attackers to generate convincing synthetic evidence is near zero, while the cost for merchants and platforms to verify the authenticity of this evidence is high, often exceeding the value of the disputed product.

2. Methodology

To investigate the mechanisms of this emerging threat, the authors conducted a qualitative study involving semi-structured interviews with two key stakeholder groups in the Chinese e-commerce ecosystem:

• Merchants (N=17): Personnel responsible for after-sales services across various storefronts (covering categories like food, cosmetics, apparel, and general goods) on platforms such as Taobao, Pinduoduo, and Douyin.
• Platform Workers (N=13): Individuals managing after-sales operations, customer service, and technical processes within the platforms (including Pinduoduo, Taobao, Meituan, and Dewu).

The study utilized a thematic analysis approach on the transcribed interview data. The authors integrated data from both groups to identify attack vectors, verification practices, and systemic constraints. The research was guided by three Research Questions (RQs):

1. What are the threat vectors in e-commerce refund fraud, and how does GenAI alter adversarial practices?
2. What verification and resolution mechanisms do merchants and platform workers currently adopt?
3. What constraints prevent the effective mitigation of these vulnerabilities?

3. Key Contributions

The paper makes three primary contributions:

1. Taxonomy of GenAI-Enabled Threat Vectors: The authors categorize refund fraud attack vectors across four stages of the transaction lifecycle: Pre-shipment/Transaction, Dispute/Evidence Submission, Logistics/Re-shipment, and Communication/Interaction.
2. Insight into Verification Vulnerabilities: The study highlights how the negligible cost of GenAI-enabled fraud shifts the burden of proof onto merchants, rendering traditional verification heuristics increasingly ineffective.
3. Design Implications for Mitigation: The paper outlines specific design implications for fraud resolution, focusing on collaborative database sharing, physical-digital grounding, and risk-adaptive adjudication workflows.

4. Key Results and Findings

4.1 Threat Vectors (RQ1)

The study identifies four distinct phases where GenAI facilitates fraud:

• Pre-shipment & Transaction: Attackers use GenAI to construct synthetic personas with realistic behavioral histories to evade detection algorithms. They also use these tools to lower the technical barrier for entry, allowing broader demographics to engage in automated promotion arbitrage and subsidy exploitation.
• Dispute & Evidence Submission: This is the primary vector. Attackers use GenAI to synthesize "hyper-realistic" defects that respect material properties (e.g., specific mold patterns on durians, snapped plastic cases). They also engage in "contextual forgery," creating fake communication records or reusing the same manipulated images across multiple shops ("digital recycling").
• Logistics & Re-shipment: Attackers exploit the time gap between automated refund approval and physical return inspection. They may use asynchronous substitution (returning counterfeit goods) or carrier interception to retain both the item and the refund.
• Communication & Interaction: Attackers use off-platform channels for "double extortion" (demanding direct transfers and then claiming a platform refund) and leverage regulatory exploitation or psychological coercion (e.g., suicide threats, regulatory complaints) to force settlements.

4.2 Verification Mechanisms (RQ2)

Merchants and platform workers employ a mix of manual and automated strategies, though these are increasingly strained:

• Manual Heuristics: Workers check for physical plausibility (e.g., unnatural fracture patterns) and request multi-angle videos or additional photos to expose spatial inconsistencies in AI generation.
• Technical Authentication: Reliance on algorithmic risk screening (detecting behavioral anomalies) and digital forensic inspection (checking Exif metadata, zooming for editing artifacts). However, these tools are often rudimentary and easily bypassed.
• Resolution Strategies: When verification fails, merchants often resort to negotiation, offering partial settlements, or absorbing the loss to avoid negative reviews and platform penalties. Escalation to human reviewers or external legal recourse is rare and often unsuccessful due to platform bias toward consumers.

4.3 Constraints on Mitigation (RQ3)

The study identifies three categories of constraints hindering effective defense:

• Technical and Infrastructural: GenAI capabilities outpace detection tools. There is a lack of dedicated, professional AI verification tools for merchants. Furthermore, the absence of "ground truth" (pre-shipment documentation, especially for perishables) makes it impossible to refute claims definitively.
• Procedural and Interactional: Platform adjudication processes are structurally biased toward consumers to ensure speed and retention. This creates an asymmetric burden of proof for merchants. Automated systems often misclassify AI-generated images as valid quality issues, and the administrative closure of disputes is often irreversible.
• Economic and Resource: The cost of producing proof (monitoring, gathering evidence) is disproportionately high for merchants compared to the low cost of attack. Additionally, the risk of reputational damage (bad reviews) forces merchants to internalize losses rather than contest claims.

5. Significance and Implications

The paper argues that the current e-commerce threat model is fundamentally broken by the advent of GenAI. The significance of this work lies in its empirical documentation of how GenAI is operationalized in real-world fraud workflows, moving beyond theoretical deepfake detection to practical economic exploitation.

The authors propose three design implications to address these vulnerabilities:

1. Cross-Platform Collaborative Defense: Implementing privacy-preserving mechanisms to share anonymized behavioral signals or hashed evidence fingerprints across platforms to identify repeat offenders and recycled evidence without violating data privacy regulations.
2. Physical-Digital Grounding: Embedding verifiable material anchors (e.g., tamper-evident markers, unique identifiers, or QR codes) into products during packaging. These anchors would serve as a ground truth that is difficult for GenAI to simulate, tying digital evidence to a specific physical transaction.
3. Risk-Adaptive Adjudication: Moving away from rigid, rule-based automation toward risk-adaptive workflows. This involves using tiered architectures where low-risk cases are automated, but high-risk or ambiguous cases are routed to specialized human oversight with dedicated, explainable AI tools to better handle complex synthetic evidence.

The paper concludes that while specific attack vectors may evolve with GenAI advancements, the underlying vulnerabilities regarding platform verification and the burden of proof will persist without structural changes to the e-commerce dispute resolution ecosystem.