PayPal
real-time fraud detection with deep learning
PayPal deployed a global real-time fraud-detection system on NVIDIA GPUs, improving detection by 10% while cutting the required server capacity by nearly 8x.
Key points
- Real-time fraud detection with deep neural networks.
- Inference moved to NVIDIA T4 GPUs, custom 24/7 scoring system.
- Detection improved by 10%, server capacity cut by nearly 8x.
- Evidence B, mixed-signals status (figures from 2019).
Objective
Detect fraud in real time on a global payments flow, 24 hours a day, without degrading checkout latency or exploding infrastructure cost.
The deployment
PayPal deployed a global real-time fraud-detection system backed by deep neural networks. CPU servers alone could not handle the load for a service that must run continuously and decide on the fly on every transaction. PayPal moved the inference to NVIDIA T4 GPUs. The result documented by NVIDIA: a 10% improvement in real-time fraud detection, while cutting the required server capacity by nearly 8x. PayPal's CTO describes a system that makes possible capabilities that were previously out of reach.
Results Proof B
Quantified deployment documented by NVIDIA (infrastructure partner, an interested source) in two consistent publications, with a direct quote from PayPal's CTO tying the brand to the project. Not a financial result; the figures come from the vendor.
How it works
Inferred typical approachThe internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.
The stack in detail
- llm Reseaux de neurones profonds PayPal PayPal's proprietary anti-fraud models, trained on the payments flow and confirmed fraud; exact architecture not published.
- infra NVIDIA T4 (GPU d'inference) GPUs to which PayPal moved real-time inference: +10% detection and server capacity cut by nearly 8x compared to CPUs.
- outil Systeme de scoring temps reel PayPal Custom chain that scores each transaction on the fly, 24/7, and routes it to approval, blocking, or human review.
How it runs, concretely
For ops teams-
1Transaction site_app
A payment is initiated on the PayPal network.
-
2Real-time scoring AI
The neural networks evaluate the transaction with low-latency GPU inference.
-
3Decision AI / risk team
The payment is approved, blocked, or put in review based on the score.
-
4Retraining data team
Confirmed fraud comes back to feed the training of the models.
Transaction behavior and confirmed fraud. If inference latency rises or the labels degrade, the system blocks legitimate payments or lets fraud through.
How your customers perceive this type of use
Sourced studiesC'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).
Acceptance conditions
- Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
- Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
- Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
- L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)
Red lines
- La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
- Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)
Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024
How to replicate
Inference, not sourcedData prerequisites
- real-time payments flow
- labeled history of confirmed fraud
- behavioral and device signals
Org prerequisites
- data science and risk engineering team
- low-latency inference infrastructure (GPU)
- governance of automated decisioning
Possible stack
- custom deep-learning model (PayPal route)
- third-party anti-fraud engines (Feedzai, Featurespace, Sift)
- GPU inference to hold latency
The plan, step by step
- Step 1Measure the baseline: latency and cost of current inference, detection rate, false-positive rate, quality of confirmed-fraud labels.Deliverable: Quantified baseline shared with the risk team
- Step 2Decide build (in-house model + GPU inference) vs a market anti-fraud engine, with a POC on replayed traffic.Deliverable: Documented detection / latency / cost comparison
- Step 3Deploy the chosen system in shadow mode on the real flow: scoring in parallel with no impact on decisions.Deliverable: Decision gaps documented over several weeks
- Step 4Gradually switch over automatic decisioning, with human review on the gray zone and dispute governance (GDPR, explainability).Deliverable: System in production with blocking and review rules
- Step 5Feed confirmed fraud back into training and monitor model drift.Deliverable: Operational label loop and drift monitoring
First step: Measure the latency and cost of current inference before deciding on a move to GPU or an external engine.
The same pattern, proven in other industries
Sources
- S1 How Is AI Used in Fraud Detection? (NVIDIA) Interested party archive pending
- S2 GPU Inference Momentum Continues to Build (NVIDIA) Interested party archive pending
- S3 Fraud Detection Applications Accelerated by NVIDIA GPUs (customer stories) Interested party archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.