Fraud remains one of the most significant challenges in the fast-paced world of eCommerce, impacting merchants and customers alike. A recent literature review featured in the Nature scientific journal showcases how a broad concensus of academic research agrees that fraud is a significant and multifaceted issue affecting various industries, from finance to healthcare and beyond.
The report offers a comprehensive overview of the most cited academia addressing the increasing sophistication of fraud schemes, including identity theft, synthetic identities, phishing, and transaction laundering. Authors throughout Nature’s review also highlight the need for continuous research and AI-driven innovations to stay ahead of evolving fraud tactics and protect businesses and customers from fraud.
Riskified, an eCommerce fraud prevention company, has also conducted advanced and probing research quantify these problems. In their survey of over 1,000 U.S. consumers, they found that 45% of respondents admitted to committing some form of “light fraud” or policy abuse. Such behavior was found to be more prevalent among younger shoppers, with 65% of those aged 18-29 admitting to such actions.
The impact of these fraudulent activities extends beyond financial losses. According to another Riskified report in partnership with Power Retail, fraud attacks have become increasingly organized and sophisticated, forcing merchants to enhance their prevention strategies or risk significant damage to both their bottom line and brand reputation
Emerj Artificial Intelligence Research recently sat down with experts from Comcast, Riskified, and eBay to talk about the challenges of managing risks by making confidence-based decisions rather than seeking absolute certainty, the use of clustering models for identifying connections between accounts in a controlled and focused manner, and the limitations of generative AI (GenAI) in providing explainable and reliable fraud detection solutions.
Riskified is a technology company that provides AI-powered fraud prevention and chargeback management solutions to eCommerce businesses, enabling them to increase sales while minimizing the risk of fraud and financial losses.
This article summarizes the below actionable insights for retail and eCommerce leaders:
- Use clustering models to detect fraud by linking suspicious accounts
- Design user experiences for legitimate users while targeting fraudsters
- Ensure transparency in credit risk models for compliance and trust
Emerj would like to thank our guests for sharing their knowledge and perspectives on fraud and risk management. In the episode summaries below, enterprise leaders can find a breakdown of these insights.
Use Clustering Models to Detect Fraud by Linking Suspicious Accounts
Episode: New Challenges in Fraud Prevention for Retail and eCommerce – with Joe Gelman of Riskified
Guest: Joe Gelman, Platform Marketing Manager at Riskified
Expertise: Branding, Product Marketing, Content Management
Brief Recognition: Joe Gelman serves as the Platform Marketing Manager at Riskified. Before this role, he was the Brand Director at Cyera. He holds a Bachelor’s Degree in Business and Philosophy from Queens College.
In the first episode in our fraud podcast series, Riskified’s Joe Gelman discusses the significant challenges merchants face in managing fraud and customer policies across the eCommerce ecosystem. He begins by highlighting the vast amount of data merchants handle, not just at checkout but throughout the entire order lifecycle. From order management and fulfillment to delivery and returns, each stage generates data that must be meticulously tracked and analyzed. Such complexity underscores the need for sophisticated tools to manage these operations effectively.
Fraud itself adds another layer of difficulty. Joe illustrates how fraudsters exploit policies by obscuring their identities using tactics like changing email addresses, device IDs, or IPs. Policies designed for individual users, such as return limits or promotional offers, are vulnerable to these manipulations.
Joe highlights one fraudster once created 365 accounts to exploit a birthday discount daily—a simple yet impactful demonstration of policy loopholes:
“They called him ‘birthday boy.’ Now, not everyone is going to be that easy to spot to have 365 accounts all back to the same person. But we have a risk, provided our engine does something called clustering. To spare you all the painful details, it basically is understanding how to tie together similar accounts. A lot of times, our dramatic analysis is picking out anomalies, things that stick out.
In this case, it’s actually figuring out who is connected to whom, and based on these similarities, you can start to group people together who want to very much not be grouped together. So it would easily group birthday boy’s activities.”
-Joe Gelman, Platform Marketing Manager at Riskified
To combat such issues, Joe discusses clustering, an advanced fraud detection method that groups similar accounts based on shared attributes or behaviors. He brings up an example of unsophisticated fraud with a merchant who had a birthday discount.
However, Joe emphasizes the importance of maintaining a balance between fraud prevention and customer experience. Overly rigid policies or false fraud detection can alienate legitimate customers, leading to lost sales and damaged trust. Merchants must ensure their systems are precise enough to prevent fraud without creating friction for genuine users.
Finally, Joe stresses the strategic importance of decisions made at critical points in the customer journey. Merchants invest heavily in acquiring customers through advertising and retargeting, and incorrect fraud detection at the point of purchase can negate these efforts. The challenge lies in making confident, data-driven decisions that protect the business while ensuring a seamless experience for legitimate customers.
He goes on to explain that risk management isn’t about eliminating risk but reaching enough confidence to make informed decisions. Absolute certainty is unattainable, and risk-taking must be strategic. He contrasts GenAI, which is less controlled and harder to understand, with their focused models like clustering, which excel at identifying connections between accounts. While less versatile, these cluster models are reliable for specific tasks, enabling adjustments as needed to improve decision-making.
Design User Experiences for Legitimate Users While Targeting Fraudsters
Episode: Navigating the New Retail and eCommerce Fraud and Risk Arms Race – with Venkatesh Palani of eBay
Guest: Venkatesh Palani, Senior Director and Head of Engineering for Risk and Trust at eBay
Expertise: Artificial Intelligence, Machine Learning
Brief Recognition: Venkatesh Palani is the Senior Director and Head of Engineering for Risk and Trust at eBay, leading a team of over 100 engineering experts. Prior to eBay, he held key roles at Mimosa Networks and Microsoft. Venkatesh holds a Master’s Degree in Engineering from the University of Melbourne.
In this podcast, Venkatesh highlights that the democratization of AI, while empowering, also equips fraudsters with advanced tools, leading to challenges like synthetic identities, deepfakes, and sophisticated phishing attacks. Traditional reliance on identity as a proxy for behavior is no longer practical, as bad actors exploit AI to create compelling content and manipulate systems, including customer support. The evolution to more advanced systems demands innovative approaches to mitigate risks in an increasingly complex environment.
Venkatesh also talks of how deepfakes undermine traditional identity verification by creating synthetic faces, making it harder to detect fraud through methods like facial matching or video comparisons, which were reliable before deepfake technology emerged. He goes on to note that with new AI-driven systems – even simple, deterministic machine learning systems – it’s much easier to see what customers value in their experience and where their attention is going:
“With the latest systems like your transformer and other advances, you’re able to see where attention should happen. So, to give you a more concrete example, you may have your traditional machine learning model, which might be used in your normal features today.
So the idea there is that if you start replacing some of these traditional features with embeddings, then what happens? You’re starting to generalize things a little bit more. You’re trying to reduce the dimensions a little bit more. You’re trying to capture more complex relationships there better.
What this does is it starts reducing the false positives, and then the system is starting to deliver a lot more accurate information.”
—Venkatesh Palani, Senior Director and Head of Engineering for Risk and Trust at eBay
He emphasizes designing user experiences tailored for legitimate users, while fraud policies focus on bad actors. A dual approach ensures minimal friction for good users while effectively addressing malicious behavior, maintaining a balance between usability and security.
In conclusion, Venkatesh explains that retailers have an advantage in combating fraud due to the vast amount of data they possess, mainly since fraudulent activity represents a much smaller portion of overall activity. The abundance of data helps retailers better understand legitimate user behavior, which is crucial as fraud tactics constantly evolve. He emphasizes the need for a balanced approach, using models that account for both good and bad behavior to avoid penalizing legitimate users.
Leveraging new technologies, like multimodal systems that analyze user interactions across various channels (e.g., text messages, communication, and site behavior), further strengthens retailers’ ability to detect and filter out fraudulent activity effectively.
Ensure Transparency in Credit Risk Models for Compliance and Trust
Episode: AI’s Role in Fraud and Credit Risk – with Shrimanth Adla of Comcast
Guest: Shrimanth Adla, Senior Director of Credit Risk at Comcast.
Expertise: Credit Risk Analytics, Artificial Intelligence
Brief Recognition: Shrimanth Adla is the Senior Director of Credit Risk at Comcast. He holds a Bachelor’s Degree in engineering from JNTU, India, and has previously worked with renowned organizations such as CitiBank, Barclays, and TransUnion.
Shrimanth’s appearance on the podcast highlights the distinct regulatory focuses on credit risk and fraud risk. Credit risk regulations prioritize financial stability and consumer protection, while fraud risk regulations aim to prevent financial crimes and identity theft. He emphasizes the importance of explainability in credit risk models, which must provide clear, detailed justifications to three key audiences: consumers (e.g., explaining why credit was denied), internal compliance teams (ensuring adherence to laws like the Fair Credit Reporting Act), and regulators (defending the model’s development and use during audits).
Ambiguity in credit risk solutions is unacceptable, as complete transparency is non-negotiable, Shrimanth explains. He goes on to note that while the fundamentals of risk management — using historical data to identify patterns and apply them to new situations — remain unchanged, the approaches differ for credit risk and fraud risk.
For credit risk, consumers may initially be legitimate but later become risky due to factors like resource shortages, allowing time to monitor and intervene with treatment plans. These factors also include tracking delinquencies over time and taking steps to manage the relationship. In contrast, fraud risk focuses on intent, with fraudsters aiming to deceive and gain access to resources. Fraud detection requires real-time or near-real-time models to quickly identify fraudulent activity, minimizing friction in the customer experience, as there’s no time to delay actions as with credit risk.
Shrimanth tells the Emerj executive podcast audience that the advancement of AI provides a more comprehensive, 360-degree view of applicants, enabling better detection of sophisticated fraud or friendly fraud. Even if two consumers have high FICO scores, alternative data and sentiment analysis can reveal which one is more likely to be fraudulent.
He also emphasizes that AI can be highly effective in preventing fraud. Still, it’s crucial to balance this sophistication with the need to minimize friction in the sales process to avoid negatively impacting the consumer experience. The key is to strike a balance between using advanced technology, adhering to regulatory requirements, and maintaining a seamless customer experience:
“I believe one should have an absolutely perfect understanding of one’s business. So that’s mandatory, and implementing the solutions is secondary because if you understand your business only, you can now identify what fits your needs and what doesn’t.
Just because you have a solution out there, jumping in and applying it to your business might not work all the time. It’s understanding your business, understanding the challenges your business has, and adapting the right technology. That’s what my advice would be.”
– Shrimanth Adla, Senior Director of Credit Risk at Comcast
