Visa AI Lead Urges Hybrid GenAI Models to Fight High-Value Fraud

Presentation by Shamindra Peiris, Visa — Ai4

Ai4

Financial institutions must route ambiguous transactions to generative models and graph neural networks while keeping traditional machine learning for instant scoring to prevent latency bottlenecks in real-time payments.

AI-assisted summary of Presentation by Shamindra Peiris, Visa — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 19, 2026, this video features a presentation by Shamindra Peiris, Senior AI Product Manager at Visa Inc. According to the page description, the talk examines how financial institutions can leverage generative AI to enhance fraud detection models, overcome severe class imbalance through synthetic data, parse unstructured data, and simulate emerging fraud scenarios.

## TL;DR Shamindra Peiris outlines how generative AI can modernize fraud detection in banking and payments to combat increasingly sophisticated, low-volume, high-value fraud schemes. The presentation details four core technical applications: generating privacy-preserving synthetic tabular data to balance training sets, utilizing Retrieval-Augmented Generation (RAG) to extract signals from unstructured text, constructing automated knowledge graphs and GNN embeddings to uncover organized fraud rings, and pairing low-latency traditional ML with GenAI copilots for ambiguous cases. Finally, Peiris highlights critical operational challenges when deploying these systems into production, including inference latency in payment authorization, model explainability, hallucinations, and emerging machine-to-machine (M2M) agentic fraud.

## Atmosphere & style The presentation carries a pragmatic, crisp, and authoritative technical tone geared toward enterprise AI practitioners and financial risk architects. Peiris delivers a tightly structured, fast-paced briefing that balances high-level industry statistics with granular architectural workflows and concrete implementation guardrails. The delivery feels purposeful and highly informative, offering an efficient, forward-looking look at modern payment security before wrapping cleanly within conference time constraints.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) Financial fraud is shifting toward low-volume, high-value schemes powered by advanced AI and deepfakes, leading to average annual losses of approximately $60 million USD per institution. - [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) Generative AI addresses severe class imbalance by generating realistic synthetic transaction data for rare fraud patterns without exposing sensitive customer PII. - [04:54](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=294s) When generating synthetic fraud datasets, practitioners should aim for a realistic 10% to 30% fraud ratio rather than an artificial 50/50 balance. - [04:54](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=294s) Validating synthetic data requires statistical alignment checks, ML fidelity evaluation against real-world holdout sets, and automated privacy audits to prevent exact record memorization. - [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) GenAI extends fraud detection beyond structured tabular data by parsing unstructured text from payment memos, customer support chats, and emails into a unified risk profile. - [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) Retrieval-Augmented Generation (RAG) combined with human-in-the-loop analyst feedback enables context-aware fraud decisions and continuous model adaptation. - [09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) GenAI automates knowledge graph construction and allows analysts to query complex fraud rings, money mule networks, and synthetic identities using natural language. - [11:51](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=711s) Graph Neural Networks (GNNs) translate graph topological relationships into numerical embeddings to empower downstream fraud classification models. - [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) A hybrid architecture leverages low-latency traditional ML for instant transaction scoring and routes ambiguous, borderline cases to GenAI copilots for deeper investigation. - [14:36](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=876s) Key production challenges for GenAI in fraud detection include real-time inference latency, hallucination risks, model explainability, and emerging machine-to-machine (M2M) agentic fraud.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) Evolving Fraud Landscape in Banking and Payments

Shamindra Peiris, Senior AI Product Manager at Visa Inc., opens his presentation at AI4 2026 by outlining the rapidly shifting dynamics of financial fraud. Recent industry data reveals that nearly 80% of businesses have reported fraud attempts, with financial institutions suffering average annual losses of approximately $60 million USD. By 2025, banking and payment fraud across North America reached critical levels, driven primarily by low-volume, high-value schemes. Rather than executing mass low-dollar scams, bad actors are focusing on extracting substantial, large sums of money from fewer individual victims.

To accomplish these high-value extractions, fraudsters are increasingly leveraging advanced technologies, including AI-enabled phishing tactics and deepfakes. Because scam methods are constantly evolving, traditional fraud detection models must evolve alongside them. Peiris notes that generative AI presents a powerful solution, outlining four primary applications that will be covered throughout the talk: synthetic data generation, unstructured data analysis, graph-based link analysis, and hybrid models combining traditional machine learning with GenAI.

- **80% of businesses** report experiencing fraud attempts. - **$60 million USD** average annual losses reported by financial institutions. - **Low-volume, high-value fraud trend** extracts massive sums from fewer targeted victims.

---

[02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) Generating Synthetic Data to Overcome Class Imbalance and PII Risks

The first primary application of GenAI in fraud analytics is generating realistic synthetic financial data to train machine learning models. A fundamental challenge in fraud detection is severe class imbalance, as legitimate transactions overwhelmingly outnumber fraudulent ones. GenAI can generate synthetic samples reflecting rare or unseen fraud patterns—such as abnormal transaction timing or new merchant category codes—without relying on real customer records.

When generating synthetic financial data, Peiris outlines a two-step process starting with identifying data types and addressing Personally Identifiable Information (PII). Fraud datasets typically consist of tabular data alongside sensitive values such as credit card numbers, full names, and tax IDs. Masking and anonymization are mandatory so models learn transaction behavioral patterns rather than specific user identities, thereby preserving data privacy.

```mermaid accTitle: Synthetic Data Generation Pipeline flowchart LR subgraph Input ["1. Data & Privacy"] A["Tabular Data & PII"] B["PII Masking & Anonymization"] A --> B end

subgraph Architecture ["2. Model Selection"] C{"Model Architecture"} D["GANs"] E["VAEs"] F["Transformer Models"] C --> D C --> E C --> F end

B --> C ```

When selecting generative architectures, practitioners generally choose between Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and transformer-based models. Peiris highlights that transformer-based models frequently outperform standard GANs when synthesizing complex, high-dimensional tabular financial data.

---

[04:54](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=294s) Best Practices for Synthetic Data Generation in Fraud Models

To ensure generated synthetic data is genuinely useful, Peiris outlines several critical best practices. First, practitioners must ensure statistical alignment by calculating high-level statistical properties on the synthetic dataset and comparing them directly against real-world distributions. Second, teams must evaluate machine learning fidelity: a baseline fraud detector (such as XGBoost or Random Forest) should be trained entirely on synthetic data and evaluated on a real-world holdout test set to verify predictive performance.

```mermaid accTitle: Best Practices for Generative Synthetic Data flowchart TD subgraph Alignment ["Validation & Fidelity"] A["Statistical Alignment Check"] B["ML Fidelity Evaluation<br>(XGBoost / Random Forest)"] end

subgraph Safeguards ["Privacy & Balance"] C["Privacy Audit Tools<br>(Prevent Record Memorization)"] D["Class Balancing<br>(Target 10% - 30% Fraud Ratio)"] E["Edge Case Simulation<br>(Rare / Unseen Fraud Schemes)"] end

A --> B B --> C C --> D D --> E ```

Additionally, organizations must perform privacy audits using specialized tooling to guarantee that the generative model does not memorize or replicate exact sensitive real-world records. Peiris cautions against fully balancing the data to a 50/50 split; maintaining a realistic class proportion of 10% to 30% fraudulent cases provides optimal training stability and model performance. Finally, generative models should intentionally simulate rare edge cases and sophisticated fraud scenarios that have not yet occurred in production, shifting detection systems from reactive to proactive defense.

- **Class balance target:** Aim for 10%–30% fraud representation rather than a 50% split. - **ML Fidelity test:** Validate models trained on synthetic data against real-world holdout data. - **Privacy protection:** Run automated privacy audits to avoid exact record memorization.

---

[07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) Harnessing Unstructured Data and RAG for Holistic Fraud Intelligence

While traditional fraud systems focus strictly on structured, tabular time-series data, modern payments generate massive volumes of unstructured data that contain critical fraud signals. GenAI enables financial institutions to parse payment memos, transaction descriptions, support emails, chat logs, and customer documentation. AI platforms can ingest both structured data and unstructured text to build a unified, holistic view of user activity across multiple channels—such as correlating phone call records, mobile application usage, and email history.

To power context-aware decision-making, organizations can implement Retrieval-Augmented Generation (RAG) to pull real-time contextual information from internal databases and policy documents. Peiris highlights the importance of integrating human-in-the-loop feedback mechanisms, where fraud analysts confirm or refute AI predictions to help the model continuously adapt to emerging fraud tactics. Furthermore, GenAI conversational assistants can summarize complex, multi-channel fraud cases for human investigators, significantly accelerating review workflows and driving operational efficiency.

- **Unstructured data sources:** Payment memos, user descriptions, support emails, chat logs, and phone call transcripts. - **Real-time enrichment:** RAG pipelines pull contextual data from internal policy documents and historical databases. - **Human feedback loop:** Fraud analysts confirm or refute predictions to continually refine model accuracy.

---

[09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) Graph-Based Link Analysis and Knowledge Graph Construction

The third key application is graph-based link analysis, where GenAI enhances graph database operations by automating query generation, deciphering complex network relationships, and discovering hidden fraud rings, synthetic identities, and money-laundering patterns. GenAI models can interact directly with graph database engines such as Amazon Neptune or Neo4j using natural language queries to uncover complex behaviors like circular transaction flows and money mule networks.

```mermaid accTitle: GenAI Graph Fraud Analytics Capabilities flowchart TD subgraph DataIngestion ["Graph Construction"] A["Unstructured Data<br>(Emails, Logs, Reports)"] B["Structured Data<br>(Transactions, IPs, Accounts)"] C["Automated Knowledge Graph<br>(Nodes & Edges)"] A & B --> C end

subgraph Analytics ["Analytics & Embeddings"] D["Natural Language Querying<br>(e.g., Neptune, Neo4j)"] E["Graph ML Embeddings<br>(GNNs capture structural context)"] F["Fraud Scheme Detection<br>(Mule networks, Bust-out fraud)"] C --> D & E E --> F end ```

GenAI automates knowledge graph construction by ingesting unstructured reports alongside structured transaction logs to construct and update graph nodes (e.g., users, physical devices, IP addresses) and edges (e.g., financial transactions, shared attributes). Analysts can query these graphs using conversational prompts—such as *"Show me all accounts sharing the same device IP in the last 24 hours"*—eliminating the need to manually write complex graph query code.

Furthermore, combining GenAI with Graph Neural Networks (GNNs) converts structural graph network patterns into numerical embeddings. These embeddings quantify how an account interacts within the wider network, capturing subtle relational links that are subsequently fed as features into downstream predictive models to uncover bust-out frauds and coordinated laundering rings.

- **Supported graph engines:** Amazon Neptune, Neo4j. - **Graph elements:** Nodes (users, devices, IPs) and Edges (transactions, shared attributes). - **Core fraud targets:** Money laundering, bust-out fraud, money mule networks, and synthetic identities.

---

[11:51](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=711s) Implementing GenAI-Powered Graph Analytics

Implementing a production-grade GenAI graph architecture requires four sequential steps: data integration, pattern definition, model training, and continuous monitoring. First, internal customer datasets must be consolidated with external sources—such as device intelligence and industry blacklists—into a unified graph database to form a comprehensive relational map.

```mermaid accTitle: Implementation Steps for Graph Analytics flowchart LR subgraph Pipeline ["Implementation Stages"] A["1. Data Integration<br>(Internal Data + Blacklists)"] B["2. Pattern Definition<br>(LLM Graph Query Logic)"] C["3. Train GNN Models<br>(DGL / AWS GraphStorm)"] D["4. Continuous Monitoring<br>(Real-time Graph Updates)"] A --> B --> C --> D end ```

Second, LLMs are used to help define specific structural query patterns, such as identifying multiple Social Security Numbers (SSNs) tied to a single individual or mapping short transaction paths to known fraud entities. Third, specialized graph frameworks like Deep Graph Library (DGL) or AWS GraphStorm are utilized to train GNN models directly on the structural topology. Finally, continuous AI monitoring updates the live graph in real time, identifying new entity relationships and instantly flagging suspicious activity.

---

[13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) Combining Traditional Machine Learning with Generative AI

Peiris emphasizes that GenAI should not replace existing machine learning infrastructure, but rather be integrated into hybrid pipelines. In an ensemble approach, traditional models like XGBoost or Random Forest handle real-time scoring due to their low latency, high throughput, and high confidence on structured data. When these traditional models encounter borderline or ambiguous cases, those specific transactions are routed to GenAI for deeper, multi-modal analysis—such as extracting context from support tickets or chat logs.

```mermaid accTitle: Hybrid Machine Learning and GenAI Workflow flowchart TD subgraph Processing ["Transaction Routing"] A["Incoming Transaction"] B["Traditional ML Model<br>(XGBoost / Random Forest)"] C["High Confidence: Instant Score"] D["Ambiguous / Doubtful Cases"] A --> B B -->|Clear Fraud / Legitimate| C B -->|Ambiguous| D end

subgraph DeepAnalysis ["GenAI Deep Analysis"] E["RAG Pipeline<br>(Pinecone / Amazon Aurora)"] F["GenAI Investigator / Copilot"] G["Automated Investigation Report"] D --> E E --> F F --> G end ```

To support deep contextual reasoning, data pipelines are augmented with RAG by connecting LLMs to vector databases such as Pinecone or Amazon Aurora. These vector stores house historical fraud cases, comprehensive customer profiles, and deep transaction logs, allowing LLMs to generate contextually precise risk assessments. Additionally, GenAI acts as an automated investigator copilot, summarizing fraud alerts, aggregating data across fragmented systems, and generating clear investigation reports to dramatically reduce manual review time.

---

[14:36](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=876s) Critical Challenges: Latency, Accuracy, Explainability, and Agentic Fraud

Peiris closes the presentation by highlighting key technical and operational challenges when deploying GenAI in production fraud systems:

- **Inference Latency:** Traditional machine learning models output predictions within microseconds, whereas GenAI models introduce higher latency. Integrating GenAI directly into real-time payment authorization flows can create significant performance bottlenecks. - **Accuracy and Hallucinations:** GenAI models risk hallucinating false details. Utilizing RAG to ground model outputs in verifiable real-time data is critical to minimizing false positive rates. - **Explainability Constraints:** Explaining traditional ML outputs is straightforward, but interpreting hybrid systems where GenAI outputs feed into machine learning models creates ambiguity. Explainable AI (XAI) frameworks must be applied to trace how LLMs arrive at specific risk scores. - **Machine-to-Machine (M2M) Agentic Fraud:** As businesses adopt autonomous agentic AI to execute routine business transactions, a new class of autonomous M2M fraud is emerging. Malicious AI agents interact directly with legitimate agents, creating severe trust and liability challenges regarding AI model outputs.

Wrapping up his session, Peiris recaps the four core applications—synthetic data generation, unstructured data analysis, graph-based link analysis, and hybrid traditional ML/GenAI integration—before thanking the organizers as the presentation reaches its time limit.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) nearly 80% of businesses have reported fraud attempts, with financial institutions suffering average annual losses of approximately $60 million USD. — Shamindra Peiris

> [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) Rather than executing mass low-dollar scams, bad actors are focusing on extracting substantial, large sums of money from fewer individual victims. — Shamindra Peiris

> [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) transformer-based models frequently outperform standard GANs when synthesizing complex, high-dimensional tabular financial data. — Shamindra Peiris

> [04:54](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=294s) maintaining a realistic class proportion of 10% to 30% fraudulent cases provides optimal training stability and model performance. — Shamindra Peiris

> [04:54](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=294s) generative models should intentionally simulate rare edge cases and sophisticated fraud scenarios that have not yet occurred in production, shifting detection systems from reactive to proactive defense. — Shamindra Peiris

> [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) GenAI conversational assistants can summarize complex, multi-channel fraud cases for human investigators, significantly accelerating review workflows and driving operational efficiency. — Shamindra Peiris

> [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) GenAI should not replace existing machine learning infrastructure, but rather be integrated into hybrid pipelines. — Shamindra Peiris

> [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) When these traditional models encounter borderline or ambiguous cases, those specific transactions are routed to GenAI for deeper, multi-modal analysis — Shamindra Peiris

> [14:36](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=876s) Traditional machine learning models output predictions within microseconds, whereas GenAI models introduce higher latency. Integrating GenAI directly into real-time payment authorization flows can create significant performance bottlenecks. — Shamindra Peiris

> [14:36](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=876s) As businesses adopt autonomous agentic AI to execute routine business transactions, a new class of autonomous M2M fraud is emerging. Malicious AI agents interact directly with legitimate agents, creating severe trust and liability challenges regarding AI model outputs. — Shamindra Peiris

## Mentioned Shamindra Peiris, Visa Inc., AI4, GANs, VAEs, Transformers, XGBoost, Random Forest, Amazon Neptune, Neo4j, Deep Graph Library (DGL), AWS GraphStorm, Pinecone, Amazon Aurora

## Video description Shamindra Peiris - Senior AI Product Manager, Visa Inc.

This talk will discuss how to leverage generative AI techniques to enhance fraud detection models.

Main challenge faced by Fraud Detection models is the class imbalance where legitimate transactions vastly outnumber fraudulent ones. Creating synthetic data will accommodate robust testing and modeling while minimizing Data Privacy concerns.

Secondly, this talk will discuss how to incorporate generative AI into the machine learning strategy to introduce novel fraud detection methods and improve predictive accuracy.

Integrating generative AI into ML strategy will shift the focus from static, rule-based detection to generating behavioral intelligence. Unlike traditional systems, generative models learn and adapt in real-time, evolving to recognize new fraud types as they emerge without requiring manual model updates.

Further, it will contribute in enhancing accuracy of the model while limiting the False Positives. At the same time, Generative AI techniques can be used to analyze unstructured data such as texts, notes etc rather than limiting to structured data points.

Lastly, this talk will discuss how to utilize generative AI to simulate potential fraud scenarios, aiding in the development and refinement of fraud prevention solutions.

[Source](https://www.youtube.com/watch?v=fFZ5fJLS_uA)

中文简报

Visa AI负责人呼吁采用混合生成式AI模型打击高价值欺诈 金融机构须将存疑交易路由至生成式模型与图神经网络，同时保留传统机器学习进行即时评分，以防实时支付出现延迟瓶颈。 Shamindra Peiris 概述了生成式 AI 如何推动银行与支付欺诈检测的现代化，以应对日益复杂、发生率低但涉案金额高的欺诈手段。演讲详述了四大核心技术应用：生成保护隐私的合成表格数据以平衡训练集、利用检索增强生成（RAG）从非结构化文本中提取特征信号、构建自动化知识图谱与 GNN 嵌入以揭露有组织的欺诈团伙，以及针对存疑交易将低延迟的传统机器学习与 GenAI 副驾驶（copilots）相结合。最后，Peiris 强调了将这些系统投入生产环境时面临的关键运营挑战，包括支付授权中的推理延迟、模型可解释性、幻觉以及新兴的机器对机器（M2M）智能体欺诈。 金融欺诈正演变为借助AI与深度伪造技术的“低频高额”模式，金融机构平均年损失达6,000万美元，促使反欺诈架构向生成式AI演进。 (0:04) 生成式AI通过构建高保真合成数据化解欺诈样本严重匮乏的痛点；实务中需严格脱敏PII并进行隐私审计，且欺诈样本比例建议维持在10%至30%。 (2:03) 结合RAG与分析师人工反馈机制，GenAI能够高效解析支付备注、客服记录等非结构化数据，实现跨渠道的全面风险洞察与模型持续演进。 (7:23) Visa建议采用混合架构部署，由XGBoost等传统机器学习模型负责毫秒级实时初筛，并将存疑交易分流至GenAI与知识图谱进行多模态深度研判。 (13:02)

Reading summary

Shamindra Peiris of Visa argues that financial institutions must augment traditional machine learning with generative AI to counter increasingly sophisticated, high-value fraud schemes. Rather than replacing existing detection infrastructure, generative tools should be deployed across synthetic data creation, unstructured data intelligence, graph analytics, and hybrid decision-making pipelines. High-Value Attack Trends and Synthetic Data Modeling (0:04) Peiris notes that financial fraud has evolved toward low-volume, high-value schemes where attackers use deepfakes and AI-enabled phishing to extract substantial sums from targeted victims. Industry data cited in the presentation indicates that nearly 80% of businesses experience fraud attempts, and financial institutions face average annual losses of roughly $60 million USD. Traditional detection systems struggle because legitimate transactions overwhelmingly outnumber fraudulent events, producing severe class imbalance that limits model training on novel attack patterns. To overcome these data shortages without compromising customer privacy, generative AI can synthesize realistic tabular financial data representing rare behaviors, unusual transaction timings, or unseen merchant category codes. Peiris emphasizes a two-step framework: identifying data types while masking Personally Identifiable Information (PII) such as credit card numbers and tax identifiers, followed by model selection. While Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are viable architectures, Peiris highlights that transformer-based models frequently outperform standard GANs when generating complex, high-dimensional tabular records. Best Practices for Synthetic Dataset Governance (4:54) Deploying synthetic data safely requires rigorous validation to ensure utility while avoiding privacy hazards. Peiris outlines two primary evaluation criteria: statistical alignment, which verifies that high-level synthetic distributions match real-world data, and machine learning fidelity, where baseline models like XGBoost or Random Forest are trained on synthetic records and tested on real-world holdout datasets. Furthermore, institutions must run automated privacy audits to guarantee that models do not memorize or replicate actual sensitive customer records. Peiris specifically cautions against balancing datasets to an artificial 50/50 ratio, recommending instead a realistic fraudulent representation of 10% to 30% to maintain training stability and detection accuracy. Generative models should also intentionally simulate rare edge cases and emerging fraud topologies that have not yet manifested in production. This practice shifts fraud defenses from reactive patching to proactive detection of unforeseen schemes. Multi-Channel Unstructured Context and Graph Link Analysis (7:23) While conventional systems rely strictly on structured tabular logs, significant fraud indicators reside in unstructured communications such as payment memos, transaction descriptions, support emails, chat transcripts, and phone logs. Implementing Retrieval-Augmented Generation (RAG) allows systems to enrich live decisions with internal policy documents and historical archives. Incorporating human-in-the-loop analyst feedback enables models to continuously adjust to shifting tactics, while conversational assistants summarize cross-channel evidence to accelerate investigations. To expose coordinated networks, GenAI automates knowledge graph construction across platforms like Amazon Neptune or Neo4j by extracting entities—such as users, IP addresses, and devices—and transaction edges from both structured logs and unstructured text. Analysts can query relational structures using natural language instead of writing complex graph queries. When integrated with Graph Neural Networks (GNNs) built via frameworks like Deep Graph Library (DGL) or AWS GraphStorm, structural graph topologies are converted into numerical embeddings that help uncover money laundering rings, bust-out fraud, and money mule networks. Hybrid Model Integration and Emerging Operational Bottlenecks (13:02) Rather than replacing legacy infrastructure, Peiris advocates for a two-tier hybrid architecture. Traditional models like XGBoost or Random Forest process incoming transactions at microsecond latency and high throughput for high-confidence decisions. Borderline or ambiguous cases are routed to generative models connected via RAG to vector databases such as Pinecone or Amazon Aurora, where they undergo multi-modal review and automated investigative summarization. However, Peiris identifies key technical and operational hurdles in production deployments. High inference latency makes GenAI unsuitable for real-time payment authorization flows without causing bottlenecks. In addition, organizations must mitigate hallucination risks through RAG grounding and deploy Explainable AI (XAI) frameworks to clarify ambiguous hybrid scoring. Finally, the rise of autonomous machine-to-machine (M2M) agentic fraud—where rogue AI agents execute automated transactions—presents novel trust and liability questions for enterprise security. Peiris concludes that effectively combating modern payment fraud requires a disciplined synthesis of low-latency traditional machine learning and contextual generative AI. However, significant unresolved challenges remain regarding the inference latency of LLMs in live payment flows, the interpretability of hybrid model outputs, and emerging security threats posed by autonomous agentic transactions.

精炼阅读摘要

Visa高级人工智能产品经理Shamindra Peiris指出，面对由深度伪造和AI网络钓鱼驱动的“低频高额”金融欺诈威胁，金融机构应通过生成式AI构建涵盖合成数据生成、非结构化分析、知识图谱与混合架构的下一代风控体系。 高额欺诈威胁演进与传统风控缺口 (0:04) 当前银行业与支付领域的欺诈格局正在快速变化。行业数据显示，接近80%的企业曾报告遭遇欺诈企图，金融机构平均每年因此蒙受约6000万美元的损失。到2025年，北美地区的金融欺诈呈现显著的“低频、高额”特征，恶意攻击者不再依赖大规模的微额欺诈，而是借助AI赋能的网络钓鱼手段及深度伪造（Deepfakes）技术，专门针对少数目标窃取大额资金。这种精细化攻击方式使仅依赖历史静态特征的传统风控模型面临失效风险，迫使机构引入生成式AI以实现主动防御。 合成数据生成与隐私保障规范 (2:03) 由于合法交易数量远超欺诈交易，欺诈检测长期受制于样本极度不平衡的问题。Peiris建议利用生成式AI合成逼真的金融数据，在不泄露客户真实信息的前提下模拟罕见的时间异常或新型商户类别代码。在技术路线上，Transformer架构在合成高维表格金融数据时的表现往往优于标准生成对抗网络（GAN）。为保护数据隐私，机构必须在建模前识别并脱敏姓名、信用卡号及税号等个人身份信息（PII），使模型专注于学习交易行为模式而非特定用户身份。 在生成合成数据时，团队需要遵循严格的最佳实践：首先，对比合成数据与真实分布的统计属性，确保统计一致性；其次，采用XGBoost或随机森林等基准分类器在合成集上训练，并在真实留出测试集上验证预测保真度。此外，机构须借助自动化隐私审计工具防止模型完全死记硬背真实记录。Peiris特别提醒，不宜将欺诈样本盲目扩充至50%的对半比例，保持10%至30%的欺诈占比能取得最佳的训练稳定性和泛化性能，同时应主动合成未曾在生产环境中出现过的极端边缘案例。 非结构化多模态整合与RAG实时增强 (7:23) 传统风控系统局限于结构化表格与时序数据，而现代支付环境产生了庞大的非结构化文本，其中蕴含关键的欺诈信号。生成式AI能够解析附言备忘录、交易描述、客服邮件、聊天记录及电话转录文本，将零散的多渠道行为进行关联，形成涵盖App使用和通信历史的立体用户画像。通过检索增强生成（RAG）技术，AI系统可在毫秒级决策中动态接入内部政策文档和历史案例库，为每笔存疑交易赋予实时上下文背景。 为了持续提升模型对新型欺诈策略的适应力，风控流程必须引入“人在回路”（Human-in-the-loop）反馈机制，由专业欺诈调查员对模型的判断进行确认或修正。与此同时，生成式AI还可充当智能调查助手，跨系统汇总碎片化信息并自动生成结构化的案件调查摘要，从而显著缩减人工审查耗时并提升运营效率。 图谱关联分析与复杂欺诈团伙挖掘 (9:14) 针对跨账户的洗钱网络、钱骡车队以及合成身份欺诈，生成式AI能显著提升图数据库（如Amazon Neptune和Neo4j）的构建与查询效率。模型能够直接解析自然语言指令，将其转化为底层的复杂关联查询，使分析师无需手写代码即可排查“过去24小时内共享同一设备IP的所有账户”或循环交易链路。该机制大幅降低了复杂网络拓扑的探索门槛。 在落地实践中，机构需整合内部交易数据与外部设备情报及黑名单，利用大模型辅助定义实体关系规则，并借助Deep Graph Library（DGL）或AWS GraphStorm训练图神经网络（GNN）。GNN可将网络交互结构转化为高维数值嵌入，表征特定账户在全局网络中的位置与关联强度，并将这些嵌入作为关键特征输入下游预测模型，从而精准侦测潜伏的破产欺诈（Bust-out fraud）和协同洗钱网络。 传统ML与生成式AI的混合协同架构 (13:02) Peiris强调，生成式AI不应完全取代现有的机器学习体系，而应作为混合风控流水线中的互补组件。传统模型（如XGBoost、随机森林）具备微秒级低延迟和高吞吐优势，适合在交易授权的第一道防线对绝大多数结构化交易进行快速实时评分。当传统模型遇到置信度不足的边缘或存疑交易时，系统再将这些交易分流至生成式AI模块，借助连接Pinecone或Amazon Aurora向量数据库的RAG流程提取上下文，进行多模态深度审查并输出自动化报告。 在生产环境中规模化部署生成式AI仍面临若干现实挑战：生成式AI较高的推理延迟难以直接适配极低耗时的实时支付授权流；模型幻觉风险要求必须依托RAG技术进行事实锚定；当生成式特征输入混合系统时，模型决策的可解释性（XAI）亟需增强。此外，随着企业广泛采用自主智能体处理业务，恶意的机器对机器（M2M）自主欺诈正在出现，对AI输出的信任度与责任归属提出了全新的安全命题。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 19 日发布在 YouTube 平台上，记录了 Visa 高级 AI 产品经理 Shamindra Peiris 的专题演讲。视频简介指出，演讲聚焦于如何利用生成式 AI 技术增强欺诈检测模型，包括解决样本极度不平衡问题、融合非结构化与图数据、构建混合架构，以及模拟欺诈场景以实现自适应防御。 ## TL;DR Visa 高级 AI 产品经理 Shamindra Peiris 在本次演讲中深入探讨了生成式 AI 赋能金融反欺诈的完整技术路径。针对银行业面临的样本不平衡、复杂团伙犯罪及多模态数据利用不足等痛点，他系统阐述了合成数据生成最佳实践、非结构化数据挖掘与 RAG 应用、图神经网络链路分析，以及传统 ML 与 GenAI 的混合集成架构。同时，他指出落地过程中必须应对时延、幻觉、可解释性以及新型智能体对智能体（M2M）欺诈等关键技术挑战。 ## Atmosphere & style 整场演讲呈现出极具专业度与严谨性的企业技术分享氛围，节奏紧凑且干货密集。主讲人以沉稳清晰的工程化视角切入，没有宽泛的行业概念炒作，而是逐一拆解架构图、数据流和落地难点，给人一种务实、系统且高度贴近金融一线风控实战的直观感受。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) 欺诈形势日益严峻，近 80% 的企业曾遭受欺诈攻击，受害机构平均年损失约 6000 万美元，且呈现出利用 AI 钓鱼与 Deepfake 进行“低频次、高金额”攻击的趋势。 - [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) 生成式 AI（GenAI）在反欺诈领域的四大应用支柱为：合成数据生成、非结构化数据分析、基于图谱的链路分析，以及传统 ML 与 GenAI 的混合集成架构。 - [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 针对反欺诈建模中严重的类别不平衡问题，GenAI 可生成高逼真度合成欺诈数据，且基于 Transformer 架构的模型在许多场景下表现已优于传统的 GAN。 - [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 合成数据生成须严格执行 PII 脱敏与掩码，且实践中欺诈样本比例应控制在 10% 至 30% 之间，避免盲目提升至 50% 造成过度平衡而损害分类器性能。 - [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) 非结构化数据分析通过整合付款备注、客服聊天、邮件及多模态数据，结合 RAG 技术挂载内部风控知识，大幅扩展了传统风控的特征维度与上下文感知能力。 - [09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) GenAI 可自动从多源日志抽取实体与行为构建知识图谱，并支持自然语言查询（NLQ），极大降低了复杂图关系排查的门槛与耗时。 - [09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) 图神经网络（GNN）结合 GenAI 生成的拓扑嵌入，能够有效识别资金循环转账、资金骡子以及团伙爆卡欺诈等高度协同的黑产行为。 - [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) 传统 ML 与 GenAI 的混合集成（Ensemble）架构采用分层漏斗模式：由 XGBoost 等传统模型负责微秒级实时初筛，存疑灰色案例则路由至 GenAI 引擎进行深度语义与多模态分析。 - [14:10](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=850s) GenAI 可作为“虚拟分析师”自动聚合欺诈告警、提炼案件摘要并生成调查报告，大幅提升反欺诈运营团队的工作流效率与案件吞吐量。 - [14:38](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=878s) GenAI 落地风控面临四大核心挑战：大模型推理带来的时延瓶颈、内容幻觉与误报风险、复杂特征链路导致的可解释性难题，以及 Agentic AI 普及催生的机器对机器（M2M）欺诈新威胁。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) 银行业欺诈形势的演变与 GenAI 的应用全景 Visa 高级产品经理 Shamindra Peiris 在 AI4 大会上发表演讲，深入探讨生成式人工智能（GenAI）在银行、金融和支付反欺诈分析中的关键应用。当前的欺诈形势正经历剧烈演变，诈骗手段日益高明。数据显示，近 80% 的企业曾遭遇欺诈攻击，受害机构在过去一年中平均损失高达约 6000 万美元。特别是在 2025 年，北美地区的银行与支付欺诈呈现出“低频次、高金额”（low-volume, high-value）的危险趋势，即犯罪分子通过锁定极少数受害者便能掠夺巨额资金。 与此类损失相对应的是欺诈工具的技术化升级：诈骗分子正广泛运用 AI 驱动的网络钓鱼（Phishing）与深度伪造（Deepfake）技术。“诈骗分子在不断进化，他们用来欺骗受害者的方法也在不断升级，因此我们的反欺诈模型也必须随之进化。” 为应对这一挑战，Peiris 梳理了 GenAI 赋能反欺诈的四大支柱：生成合成数据以解决样本匮乏、分析非结构化数据、基于图数据库的关联链路分析（Graph-based link analysis），以及将传统机器学习（ML）与 GenAI 深度融合。 - **欺诈损失现状**：近 80% 企业遭遇欺诈，机构平均年损失约 6000 万美元。 - **作案手段趋势**：利用 AI 钓鱼与 Deepfake 进行精准定向的大额欺诈。 - **应对核心路径**：合成数据生成、非结构化数据分析、知识图谱链路挖掘、传统 ML 与 GenAI 协同。 --- ### [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 基于 GenAI 的合成数据生成及最佳实践 反欺诈建模面临的最大痛点是极端的“类别不平衡”（Class imbalance）——正常交易在体量上占据绝对主导，而欺诈样本极其稀少。GenAI 能够生成高度逼真的合成金融数据，用于训练机器学习模型识别如异常交易时间或新型商户类别等未知欺诈模式。在实施过程中，团队首先需明确数据形态（以表格型交易数据为主），并严格执行 PII（个人身份信息）脱敏与掩码（PII Masking），确保模型仅学习交易行为模式而非特定个人身份，从根源上保护数据隐私。 在模型架构选型上，业界常采用生成对抗网络（GAN）、变分自编码器（VAE）以及 Transformer 架构模型。Peiris 指出，基于 Transformer 的模型在许多场景下的表现已超越传统的 GAN。 ```mermaid flowchart LR accTitle: 合成数据生成流程 subgraph DataPrep ["数据准备与隐私保护"] A["真实交易数据<br/>(表格/时序)"] --> B["PII 识别与脱敏掩码"] end subgraph ModelGen ["生成模型建模"] B --> C{"生成模型选型<br/>(Transformer / GAN / VAE)"} C --> D["生成合成欺诈数据"] end subgraph EvalDeploy ["验证与下游应用"] D --> E["统计一致性与保真度评估"] E --> F["训练下游检测器<br/>(XGBoost / Random Forest)"] end ``` 在生成合成数据时，必须遵循一系列工程最佳实践： 1. **统计对齐（Statistical Alignment）**：计算关键统计指标，验证合成数据与真实基准数据在宏观分布上的一致性。 2. **机器学习保真度（ML Fidelity）**：在合成数据集上训练反欺诈分类器（如 XGBoost、Random Forest），并在独立的真实保留数据集（Holdout dataset）上评估其表现。 3. **隐私审计（Privacy Audit）**：使用开源审计工具进行检测，严防生成模型“死记硬背”并输出敏感的真实世界记录。 4. **控制样本比例（避免过度平衡）**：切忌将欺诈比例盲目提升至 50%；实践表明，将合成欺诈样本比例控制在 10% 至 30% 最有利于下游分类器发挥最佳性能。 5. **聚焦边缘案例（Edge Cases）**：利用生成模型针对性模拟罕见且复杂的潜在欺诈场景，使防御机制从被动响应转向主动预防。 --- ### [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) 非结构化数据分析与多模态反欺诈 传统的反欺诈系统多局限于表格数据与时间序列指标，但在现代支付生态中，隐藏在非结构化数据中的欺诈线索愈发关键。这涵盖了付款备注（Payment memos）、交易描述、电子邮件、客服聊天记录（Chat logs）及各类支持文档。 通过 AI 平台整合多渠道数据源，风控系统能够将电话呼叫记录、应用使用轨迹与电子邮件等非结构化线索与底层结构化交易合并，构建全方位的用户行为画像。在此架构中，检索增强生成（RAG）技术发挥着核心作用：通过从内部数据库和政策文档中实时检索背景信息，大模型能够在充分掌握上下文的情况下做出风险决策。同时，构建包含人工调查员（Fraud Analysts）的闭环反馈机制至关重要，调查员对预测结果的确认或驳回将直接驱动模型持续迭代。此外，基于 GenAI 的智能助手可自动提炼复杂案件摘要，大幅提升一线运营人员的调查效率。 - **多模态数据摄入**：整合通话、App 行为、邮件与交易备注等多维度数据。 - **RAG 驱动风控**：实时挂载内部知识库与风控策略，确保决策具备上下文感知能力。 - **人机协同反馈**：调查员确认/否定预测结果，反哺模型自适应新型攻击手段。 --- ### [09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) 基于图谱的链路分析（Graph-Based Link Analysis） 面对洗钱、团伙欺诈（Fraud rings）及合成身份（Synthetic identities）等复杂协同犯罪，GenAI 与图技术的结合开辟了新的防御维度。GenAI 能够从非结构化报告与结构化流水中自动抽取实体（用户、设备、IP）作为图节点（Nodes），抽取交易与关联行为作为边（Edges），全自动构建并更新反欺诈知识图谱。 ```mermaid flowchart TD accTitle: GenAI 驱动的图分析工作流 subgraph Ingestion ["数据摄入与图谱构建"] G1["多源数据<br/>(报告/日志/设备黑名单)"] --> G2["自动化图谱构建<br/>(节点: 用户/IP; 边: 交易)"] end subgraph GraphProc ["图计算与特征工程"] G2 --> G3["NLQ 交互查询<br/>(Neptune / Neo4j)"] G2 --> G4["图神经网络 (GNN)<br/>数值嵌入特征提取"] end subgraph Detection ["欺诈团伙识别"] G3 --> G5["实时拦截欺诈团伙 & 骡子账户"] G4 --> G5 end ``` 在交互层面，GenAI 赋予了图数据库（如 Amazon Neptune、Neo4j）强大的自然语言查询（NLQ）能力。调查人员无需编写复杂的图查询语句，仅需输入类似 "Show me all the accounts sharing the same device IP in the last 24 hours"（显示过去 24 小时内共享同一设备 IP 的所有账户）等自然语言提示词，即可实现毫秒级排查。 在底层算法层面，图神经网络（GNN）配合 GenAI 将高维图结构信息转化为数值嵌入（Embeddings），捕捉账户在整个网络中的深层拓扑关系，进而识别出资金循环闭环转账（Circular transactions）、资金骡子（Mule behaviors）以及爆卡欺诈（Bust-out frauds）。落地实施该方案需依托系统化工程： - **全量数据融合**：整合内部客户档案与外部设备威胁情报、黑名单。 - **LLM 辅助规则定义**：利用大模型辅助定义图谱查询模式（如单人关联多个社保号 SSN，或与已知欺诈者存在超短跳转路径）。 - **专业框架训练**：使用 Deep Graph Library (DGL) 或 AWS GraphStorm 等分布式框架训练图模型。 - **实时图谱更新**：持续动态维护拓扑网络并对异常拓扑聚集进行秒级告警。 --- ### [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) 传统 ML 与 GenAI 的混合集成架构 为平衡高并发实时性与深层次语义推理，产业界正加速构建传统机器学习与 GenAI 的混合集成（Ensemble）架构。在典型的风控流水线中，传统模型与生成式模型各司其职，形成分层漏斗： - **分流机制**：由 XGBoost、Random Forest 等传统高效算法承担第一道防线，进行高吞吐、毫秒级的高置信度实时评分（High-confidence real-time scoring）。 - **复杂判障上报**：对于处于灰色地带的模糊、存疑交易案例（Ambiguous cases），系统将其自动路由至 GenAI 引擎，深入解析关联聊天记录、客服工单等非结构化数据并生成归因洞察。 - **RAG 数据增强**：利用向量数据库（如 Pinecone、Amazon Aurora）连接大语言模型，实时召回历史欺诈档案、用户画像与历史交易上下文，为复杂风险评估提供精准的情境支持。 ### [14:10](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=850s) 利用生成式 AI 自动化欺诈调查工作流程 生成式 AI（GenAI）在反欺诈领域的另一个关键落地场景是自动化欺诈调查工作流程。在实际风控运营中，企业完全可以将生成式 AI 部署为“虚拟分析师”，由其负责汇总分析各类欺诈告警、综合并提炼来自不同数据源的多维信息，并最终自动生成结构化的调查报告。这种智能化的调查辅助手段能够极大程度削减人工复核的时间成本，显著提升整个反欺诈团队的运营效率。 ```mermaid flowchart LR accTitle: GenAI 自动化欺诈调查工作流 subgraph DataInput["多源数据输入"] A["多源风控数据"] B["欺诈告警信号"] end subgraph AgenticProcessing["GenAI 分析引擎"] C["数据综合与提炼"] D["自动化生成调查报告"] end subgraph Outcome["业务成效"] E["缩短人工复核耗时"] F["提高案件处理吞吐量"] end A --> C B --> C C --> D D --> E D --> F ``` --- ### [14:38](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=878s) 生成式 AI 落地欺诈风控面临的核心挑战与总结 在将生成式 AI 深度融入欺诈检测体系时，团队必须应对四大维度的核心技术与业务挑战： 1. **时延（Latency）瓶颈**：在生产部署环境中，传统的机器学习（ML）模型推理极为轻量高效，通常能在微秒级（microseconds）内完成评分输出。然而，一旦引入结构化多模态数据分析，或将 GenAI 大模型嵌入到传统的机器学习工作流中，端到端系统的响应延迟必然会大幅上升，这对实时交易风控构成了严峻考验。 2. **准确度与幻觉控制（Accuracy & Hallucinations）**：生成式大模型天然存在生成错误信息（幻觉）的风险。为了保证风控决策的精确并降低误报率（False Positives），必须引入检索增强生成（RAG）等技术手段，使大模型基于实时的真实世界事实数据进行逻辑推理和决策。 3. **可解释性（Explainability）难题**：传统的机器学习模型拥有相对成熟且易于理解的可解释性分析方法。然而，当大语言模型（LLM）的输出或特征被输入到传统机器学习模型中时，整个风控系统做出的风险评分归因将变得模糊而复杂。风控团队需要借助专门的可解释性 AI（XAI）框架，明确解析 LLM 是如何得出特定风险评级结论的。 4. **机器对机器（M2M）欺诈新威胁**：随着越来越多的企业采用 Agentic AI（智能体 AI）来自主处理日常交易，自主化的“机器对机器”（M2M）欺诈正呈现上升趋势。恶意 AI 智能体与合规智能体之间的交互博弈，给 AI 模型输出的信任度（Trust）与责任归属（Liability）带来了巨大的不确定性。 演讲最后总结回顾了生成式 AI 在欺诈检测中的四大核心应用方向：合成数据生成（Synthetic Data Generation）、非结构化数据分析（Unstructured Data Analysis）、图数据关联分析（Graph-based Data Analysis），以及传统机器学习与生成式 AI 模型的混合架构设计，并强调在推进技术落地的同时必须妥善解决上述高维度挑战。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) 诈骗分子在不断进化，他们用来欺骗受害者的方法也在不断升级，因此我们的反欺诈模型也必须随之进化。 — Shamindra Peiris > [00:04](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=4s) 近 80% 的企业曾遭遇欺诈攻击，受害机构在过去一年中平均损失高达约 6000 万美元。 — Shamindra Peiris > [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 反欺诈建模面临的最大痛点是极端的‘类别不平衡’——正常交易在体量上占据绝对主导，而欺诈样本极其稀少。 — Shamindra Peiris > [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 切忌将欺诈比例盲目提升至 50%；实践表明，将合成欺诈样本比例控制在 10% 至 30% 最有利于下游分类器发挥最佳性能。 — Shamindra Peiris > [02:03](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=123s) 利用生成模型针对性模拟罕见且复杂的潜在欺诈场景，使防御机制从被动响应转向主动预防。 — Shamindra Peiris > [07:23](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=443s) 通过从内部数据库和政策文档中实时检索背景信息，大模型能够在充分掌握上下文的情况下做出风险决策。 — Shamindra Peiris > [09:14](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=554s) GenAI 能够从非结构化报告与结构化流水中自动抽取实体作为图节点，抽取交易与关联行为作为边，全自动构建并更新反欺诈知识图谱。 — Shamindra Peiris > [13:02](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=782s) 由 XGBoost、Random Forest 等传统高效算法承担第一道防线，进行高吞吐、毫秒级的高置信度实时评分。 — Shamindra Peiris > [14:10](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=850s) 企业完全可以将生成式 AI 部署为‘虚拟分析师’，由其负责汇总分析各类欺诈告警、综合并提炼来自不同数据源的多维信息，并最终自动生成结构化的调查报告。 — Shamindra Peiris > [14:38](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=878s) 在生产部署环境中，传统的机器学习模型推理极为轻量高效，通常能在微秒级内完成评分输出。然而，一旦引入结构化多模态数据分析，端到端系统的响应延迟必然会大幅上升。 — Shamindra Peiris > [14:38](https://www.youtube.com/watch?v=fFZ5fJLS_uA&t=878s) 随着越来越多的企业采用 Agentic AI 来自主处理日常交易，自主化的‘机器对机器’（M2M）欺诈正呈现上升趋势。 — Shamindra Peiris ## Mentioned Shamindra Peiris, Visa, Ai4, GAN, VAE, Transformer, XGBoost, Random Forest, RAG, Amazon Neptune, Neo4j, GNN (图神经网络), Deep Graph Library (DGL), AWS GraphStorm, Pinecone, Amazon Aurora, Agentic AI, XAI (可解释性 AI) ## Video description Shamindra Peiris - Senior AI Product Manager, Visa Inc. This talk will discuss how to leverage generative AI techniques to enhance fraud detection models. Main challenge faced by Fraud Detection models is the class imbalance where legitimate transactions vastly outnumber fraudulent ones. Creating synthetic data will accommodate robust testing and modeling while minimizing Data Privacy concerns. Secondly, this talk will discuss how to incorporate generative AI into the machine learning strategy to introduce novel fraud detection methods and improve predictive accuracy. Integrating generative AI into ML strategy will shift the focus from static, rule-based detection to generating behavioral intelligence. Unlike traditional systems, generative models learn and adapt in real-time, evolving to recognize new fraud types as they emerge without requiring manual model updates. Further, it will contribute in enhancing accuracy of the model while limiting the False Positives. At the same time, Generative AI techniques can be used to analyze unstructured data such as texts, notes etc rather than limiting to structured data points. Lastly, this talk will discuss how to utilize generative AI to simulate potential fraud scenarios, aiding in the development and refinement of fraud prevention solutions. [Source](https://www.youtube.com/watch?v=fFZ5fJLS_uA)

Source quotes

[0:04] Shamindra Peiris: 诈骗分子在不断进化，他们用来欺骗受害者的方法也在不断升级，因此我们的反欺诈模型也必须随之进化。 [0:04] Shamindra Peiris: 近 80% 的企业曾遭遇欺诈攻击，受害机构在过去一年中平均损失高达约 6000 万美元。 [2:03] Shamindra Peiris: 反欺诈建模面临的最大痛点是极端的‘类别不平衡’——正常交易在体量上占据绝对主导，而欺诈样本极其稀少。 [2:03] Shamindra Peiris: 切忌将欺诈比例盲目提升至 50%；实践表明，将合成欺诈样本比例控制在 10% 至 30% 最有利于下游分类器发挥最佳性能。 [2:03] Shamindra Peiris: 利用生成模型针对性模拟罕见且复杂的潜在欺诈场景，使防御机制从被动响应转向主动预防。 [7:23] Shamindra Peiris: 通过从内部数据库和政策文档中实时检索背景信息，大模型能够在充分掌握上下文的情况下做出风险决策。 [9:14] Shamindra Peiris: GenAI 能够从非结构化报告与结构化流水中自动抽取实体作为图节点，抽取交易与关联行为作为边，全自动构建并更新反欺诈知识图谱。 [13:02] Shamindra Peiris: 由 XGBoost、Random Forest 等传统高效算法承担第一道防线，进行高吞吐、毫秒级的高置信度实时评分。 [14:10] Shamindra Peiris: 企业完全可以将生成式 AI 部署为‘虚拟分析师’，由其负责汇总分析各类欺诈告警、综合并提炼来自不同数据源的多维信息，并最终自动生成结构化的调查报告。 [14:38] Shamindra Peiris: 在生产部署环境中，传统的机器学习模型推理极为轻量高效，通常能在微秒级内完成评分输出。然而，一旦引入结构化多模态数据分析，端到端系统的响应延迟必然会大幅上升。 [14:38] Shamindra Peiris: 随着越来越多的企业采用 Agentic AI 来自主处理日常交易，自主化的‘机器对机器’（M2M）欺诈正呈现上升趋势。

Source quotes

[0:04] Shamindra Peiris: nearly 80% of businesses have reported fraud attempts, with financial institutions suffering average annual losses of approximately $60 million USD. [0:04] Shamindra Peiris: Rather than executing mass low-dollar scams, bad actors are focusing on extracting substantial, large sums of money from fewer individual victims. [2:03] Shamindra Peiris: transformer-based models frequently outperform standard GANs when synthesizing complex, high-dimensional tabular financial data. [4:54] Shamindra Peiris: maintaining a realistic class proportion of 10% to 30% fraudulent cases provides optimal training stability and model performance. [4:54] Shamindra Peiris: generative models should intentionally simulate rare edge cases and sophisticated fraud scenarios that have not yet occurred in production, shifting detection systems from reactive to proactive defense. [7:23] Shamindra Peiris: GenAI conversational assistants can summarize complex, multi-channel fraud cases for human investigators, significantly accelerating review workflows and driving operational efficiency. [13:02] Shamindra Peiris: GenAI should not replace existing machine learning infrastructure, but rather be integrated into hybrid pipelines. [13:02] Shamindra Peiris: When these traditional models encounter borderline or ambiguous cases, those specific transactions are routed to GenAI for deeper, multi-modal analysis [14:36] Shamindra Peiris: Traditional machine learning models output predictions within microseconds, whereas GenAI models introduce higher latency. Integrating GenAI directly into real-time payment authorization flows can create significant performance bottlenecks. [14:36] Shamindra Peiris: As businesses adopt autonomous agentic AI to execute routine business transactions, a new class of autonomous M2M fraud is emerging. Malicious AI agents interact directly with legitimate agents, creating severe trust and liability challenges regarding AI model outputs.

Original transcript

[0:04] Good morning, everyone. I'm Shamindra Peries  from Visa. Uh currently working as a senior [0:11] product manager. And glad to be part of AI 4  2026. And thanks for the organizing committee [0:18] for this wonderful opportunity. So today, um  I'll be speaking about the applications of GenAI [0:26] in fraud analytics in banking,  finance, and payments. Um [0:34] So let's first start with some high-level  information on um the evolving fraud landscape. [0:40] And then we can get into the details  of how GenAI can be leveraged uh [0:46] in fraud analytics. Starting with synthetic data  generation, then the unstructured data analysis, [0:52] uh then one of the most important applications,  the graph-based link analysis, and finally, [0:58] how traditional machine learning  models can be combined with GenAI. [1:05] So as you can see here, um organizations indicate  that nearly 80% of the businesses have experienced [1:13] fraud attempts with institutions reporting  average losses of approximately USD 60 million [1:19] over the past year. In 2025, banking  and payment fraud across North America [1:26] escalated to critical levels, reflecting  low-volume, high-value frauds. Um meaning like, [1:34] you know, extracting substantial large sums uh by  just, you know, like scamming just fewer victims. [1:43] So reports also suggest that the scammers  are increasingly leveraging AI-enabled [1:49] phishing tactics and deepfake technology. So  it's clear that the uh the scammers are evolving, [1:55] and the methods that they use to scam the victims  are evolving, so definitely our fraud models [2:03] need to evolve as well. So, the first application  is synthetic data generation using GenAI. [2:12] Um as you know, like you know, GenAI  creates realistic synthetic financial data [2:17] to train machine learning models, uh helping them  identify new previously unseen fraud patterns, [2:25] um such as unusual transaction times  or like new merchant categories, etc. [2:32] So, let's look at um two main  steps involved in this process. [2:38] So, as you can see here, uh the first  step is um identifying the data types. [2:45] Um as you may already know, um most of the time  the fraud data is, like you know, tabular form. [2:52] Maybe the transaction amount and other related  information. And um one of the most um like you [3:01] know, significant challenges that we face with  fraud data is the class imbalance. Because the um [3:10] the fraudulent data is, like you know, it's  it's very low compared to the the legitimate [3:16] transactions. And the next phase is handling PII,  or like the personally identifiable information. [3:23] As you know, like you know, fraud data can  um include information like the card numbers, [3:29] names, uh tax IDs, things like that.  So, it's very important that we do [3:36] the necessary um you know, PII masking uh in  in the synthetic data generation. And also, [3:44] like you know, when synthetic data being  generated, uh we can use, like you know, [3:49] certain type of anonymizations and stuff,  so that um the models will be trained [3:56] for data or the patterns and not for  the identities and at the same time [4:01] we can make sure the identities are you know, it's  it's masked and the data privacy is maintained. [4:10] Then the next step is to identify um the best  model. So, as you can see here, I have like listed [4:19] down like three models, um generative adversarial  networks, um which is like the most popular one. [4:26] And the other one is the variational autoencoders.  Um finally, the transform-based models. [4:34] Um mostly, you know, this method um kind of  outperform the GANs. If you have actually [4:41] worked with um these models in your corporate  life or like, you know, in any of the projects, [4:47] you would know like, you know, how these  models would you know vary from each other. [4:54] Now, let's look at some of the um best practices  to be used uh when using GenAI for synthetic data [5:03] generation. So, as you can see here, uh first,  I think we need to make sure like, you know, [5:12] once the synthetic data is generated, that it's  statistically aligned with your genuine data [5:18] or the real data. So, make sure that  you probably calculate some um you know [5:24] uh high-level uh stats and compare it against your  real data set to see that, you know, that you have [5:30] that alignment with the real data. And then, the  machine learning fidelity. So, make sure uh your [5:41] um like, you know, the once once the  synthetic data set is generated, you're [5:45] like train a fraud detector. for example, like  XGBoost or random forest on the synthetic data [5:51] set and test it out on the holdout data set to  see how the model performs. If you see like, [5:57] you know, the alignment there, then  the the data that you have used is [6:03] you know, you can say that the synthetic  data has been generated uh as expected. [6:10] So, yeah, we can start with the the machine  learning fidelity. So, as I mentioned earlier, [6:14] uh we could um train the fraud detector like, you  know, XGBoost or random forest for the synthetic [6:21] data and see how it performs. And then uh thirdly,  the privacy audit. Um use tools uh to ensure the [6:32] generated data does not memorize and repeat uh  sensitive real-world records. Um there are some [6:39] open-source uh tools that you could use for this.  I'm not going to mention the the brand names. [6:45] Um and also like um the other most important  thing is make sure that it does not fully balance. [6:52] It is often best not to make the fraud cases equal  to 50% of the data. Aiming for 10 to 30% would be [7:01] uh sufficient for optimal uh machine learning  performance. Finally, uh focus on the edge cases. [7:09] So, use the generative model to simulate rare,  sophisticated fraud scenarios that have not yet [7:16] occurred, allowing detection models to become  more proactive in identifying the frauds. [7:23] Now, let's move to the second application,  which is the unstructured data and see. So, [7:29] as we discussed earlier, usually the fraud data  is in a tabular form or like a time series form, [7:35] but as we know, like when it comes to payments,  um there are a lot of unstructured data that [7:40] we can use. for example, um payment  memos, descriptions, emails, chat logs, [7:46] and other documents to identify any fraud  patterns. So, let's see how we can implement um [7:54] GenAI for analyzing the unstructured data.  First and foremost, we need to find out the [8:00] way to integrate the data sources. So, we can use  certain AI platforms to ingest both structured [8:07] data and unstructured data to provide a holistic  view of user activities across the channels. Um [8:14] for example, like merging phone calls, application  usage, and emails, etc. Then, you could employ rag [8:22] uh to pull relevant real-time data from internal  databases and maybe policy documents and etc. [8:30] allowing the AI to make context-aware decisions  on fraud risks. Um then, another important thing [8:38] is to make sure that you integrate GenAI with  feedback mechanisms where investigators or like, [8:44] you know, real-life uh fraud analysts confirm or  refute the predictions. So, that it would allow [8:50] the model to continuously adapt to the new  fraud tactics. Finally, we could um leverage [8:58] GenAI chatbots to assist the fraud analysts by  summarizing complex cases and providing rapid [9:04] insights. So, this would definitely make the  entire process more uh operationally efficient. [9:14] Now, let's look at the the third application,  which is the graph-based link analysis. [9:19] Um GenAI enhances graph-based link analysis  for fraud by automating query generation, [9:27] interpreting complex network relationships, and  synthesizing insights to identify fraud rings, [9:34] uh synthetic identities, and uh  certain money laundering uh trends. [9:40] Also, it enables the real-time natural language  querying of graph databases. For example, [9:45] Neptune or Neo4j to uncover hidden patterns  like circular transactions or mule behaviors. Um [9:55] So, now let's look at, you know, some of the key  applications of uh GenAI in graph fraud analytics. [10:03] Um first one being the automated  uh knowledge graph construction. [10:09] So, GenAI can ingest unstructured data such  as reports, emails, chat logs, and structured [10:14] data to build um build and map and update graph  nodes such as entities like users, devices, IPs, [10:21] and edges such as transactions and relationships  automatically. So, if you have actually worked [10:28] with the um graph-based models, you would know  it actually works with nodes and edges. So, it's [10:34] important that how you define the edges and nodes  uh when defining this graph-based. And also, um [10:41] secondly, the natural language query generation.  Uh fraud analysts can use natural language. [10:47] Uh for example, queries like, "Show me all  the accounts sharing the same device IP in [10:53] the last 24 hours or last 12 hours." Um So,  this will definitely, you know, like accelerate [11:00] the investigations and it will definitely  make the process more efficient. Then, um [11:06] graph ML feature engineering and enrichment.  GenAI and graph neural networks convert structural [11:13] graph data into numerical embeddings. These  embeddings can capture how an account interacts [11:19] within the network, identifying hidden links that  are then fed into the predictive fraud models. [11:27] Um Finally, like, you know, visualizing  and detecting the fraud rings. [11:32] GenAI aids in detecting complex  schemes such as money laundering, [11:37] uh bust-out frauds, and synthetic identities  by identifying the clusters or mule behaviors. [11:44] Uh and certain, you know, like money laundering  patterns are also can be identified through this. [11:51] Now, let's see how to implement  GenAI-powered uh graph analytics. [11:57] Um first one is, you know, the data integration.  Um consolidate internal, for example, like, [12:04] you know, customer data and external  data, like, you know, device intelligence, [12:09] uh certain blacklists into the graph database  to build a comprehensive relation map. Then, [12:15] uh defining the fraud patterns using LLMs to  help define graph query patterns. For example, [12:23] uh multiple SSNs used by one person or a short  path known to a fraudster, something like that. [12:30] And then, uh training and deploying  the GNN models. Use frameworks like uh [12:38] Deep Graph Library and AWS um GraphStorm to train  models on the graph structure. Finally, it's very [12:46] important that you do the continuous monitoring.  So, use AI constantly to update the graph, [12:52] identify the new relationships, and flagging  suspicious active uh activities immediately. [13:02] So, the final application is where you  can combine the traditional ML with GenAI. [13:10] Um let's see how we can integrate  GenAI into existing pipelines [13:14] in the different steps of the machine learning  uh model cycle. So, implement an ensemble models. [13:21] Use existing traditional models, for  example, XGBoost, random forest forest uh [13:26] for high-confidence real-time scoring. and route  ambiguous cases with doubtful cases to GenAI. [13:36] For deeper analysis and for, you know, like um  maybe to extract the unstructured data and um [13:43] come up with certain insights uh through  chat logs or any other support tickets. [13:50] Uh then enhance the data pipelines with RAG. So,  use RAG by connecting LLMs to vector databases. [13:57] Um maybe Pinecone, Amazon Aurora. Um So,  which contains historical fraud cases, [14:04] customer profiles, and transaction logs.  So, this allows LLMs to generate contextual [14:10] accurate fraud assessments. Finally, automate  investigation workflows. So, we can definitely [14:18] deploy GenAI as an analyst uh to summarize fraud  alerts, synthesize data from different sources, [14:26] and generate investigation reports. So, this  will definitely uh reduce the manual review time [14:32] and make the process more efficient. So, finally,  I have uh listed down some other challenges with [14:38] GenAI. Um the latency. Because if it is like a  traditional machine learning model, usually um [14:47] at the deployment level or in production,  the latency time is like, you know, [14:51] it's very minimal. Within like microseconds,  you can generate an output. [clears throat] But [14:55] if you're using structured data or any GenAI  models embedded in your traditional machine [15:01] learning model, it will definitely, you  know, like uh increases the latency. [15:06] Then about the accuracy. Um you have  hallucinations and stuff, and um [15:12] it's very important that we use RAG to  ground the GenAI in factual real-time data, [15:17] reducing the false positives. Uh finally, um the  explainability. So use explainable AI frameworks [15:28] to understand how the LLM arrived at a specific  risk rating. In traditional machine learning, [15:33] it's very easy to kind of you know, like  explain the output of a machine learning, [15:38] but when the GenAI is output is kind of input  into your traditional machine learning model, [15:44] it's it's become a bit ambiguous how you're  going to like, you know, interpret that output. [15:51] Fourth one is the machine to machine scams, [15:54] which is another trend that we could see. So as  businesses adopt agentic AI to handle routine [15:59] transaction, a rise in autonomous M2M fraud where  malicious agents interact with legitimate ones, [16:07] creating massive uncertainty regarding the the  trust and the the liability of the output of the [16:17] output of your AI model. So we went  through like four applications, [16:24] synthetic data generation, unstructured  data analysis, graph-based data analysis, [16:30] and then finally combining traditional machine  learning and GenAI models and talk about some [16:37] of the high-level challenges as well. So that's  the end of my presentation and we are at time.