AI-Powered Fraud Detection Using Mixture of Experts and Federated Learning

feedzai · Finance · 2025 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Financial crime and fraud detection

Feedzai needed fraud detection that could adapt to emerging cross-institution fraud patterns without requiring every financial institution to wait months for isolated custom models.

Constraints:

Fraud patterns are adversarial and evolve across institutions and geographies
Financial institutions cannot share raw transaction data freely
Risk scores must remain stable, interpretable, and compatible with downstream systems

Architecture

Feedzai TrustScore is a fraud-risk scoring architecture built from pre-trained expert models. It standardizes institution data into a shared schema, assembles relevant experts through a Mixture of Experts design, aggregates cross-institution fraud knowledge through federated learning without moving client data, calibrates expert scores into a stable risk distribution, and keeps the system refreshed and monitored as fraud patterns evolve.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  institution-event-schema["institution-event-schema"]
  data-harmonization-pipeline["data-harmonization-pipeline"]
  pretrained-fraud-experts["pretrained-fraud-experts"]
  federated-learning-layer["federated-learning-layer"]
  moe-expert-panel["moe-expert-panel"]
  expert-calibration-layer["expert-calibration-layer"]
  score-monitoring-loop["score-monitoring-loop"]
  risk-score-output["risk-score-output"]
  client-fraud-stack["client-fraud-stack"]
  institution-event-schema --> data-harmonization-pipeline
  data-harmonization-pipeline --> moe-expert-panel
  pretrained-fraud-experts --> federated-learning-layer
  federated-learning-layer --> moe-expert-panel
  moe-expert-panel --> expert-calibration-layer
  expert-calibration-layer --> risk-score-output
  expert-calibration-layer --> score-monitoring-loop
  score-monitoring-loop --> pretrained-fraud-experts
  risk-score-output --> client-fraud-stack

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`institution-event-schema`	standardizes tabular financial events so expert patterns can transfer across institutions	`data-harmonization-pipeline`, `moe-expert-panel`
`data-harmonization-pipeline`	performs schema alignment, categorical harmonization, currency conversion, and time-zone feature derivation	`institution-event-schema`, `moe-expert-panel`
`pretrained-fraud-experts`	encodes learned fraud patterns from specific geographies and use cases	`moe-expert-panel`, `federated-learning-layer`
`federated-learning-layer`	aggregates fraud knowledge as model encodings while keeping raw customer data in secure environments	`pretrained-fraud-experts`, `moe-expert-panel`
`moe-expert-panel`	combines relevant expert models to produce risk decisions for a client environment	`expert-calibration-layer`, `risk-score-output`
`expert-calibration-layer`	normalizes expert contributions and keeps scores comparable over time	`moe-expert-panel`, `risk-score-output`, `score-monitoring-loop`
`score-monitoring-loop`	monitors expert and aggregate performance, drift, and when to add, remove, or update experts	`pretrained-fraud-experts`, `expert-calibration-layer`
`risk-score-output`	provides a stable fraud-risk signal that can augment rules, custom AI models, and downstream reviews	`client-fraud-stack`
`client-fraud-stack`	combines TrustScore with existing rules or custom AI workflows	`risk-score-output`

Data flow

Financial-institution events are aligned into a standard schema with harmonized categorical fields, currencies, and derived time-zone features. Relevant pre-trained expert models, each specialized by geography or use case, are combined as a panel of specialists. Federated learning shares high-level model encodings rather than customer data. Expert and MoE-level calibration produce a consistent risk score that downstream fraud workflows can combine with rules or custom models.

3 supporting quotes

“To develop a unified AI solution, applicable to multiple institutions, the format and the meaning of the data must be standardized”article, stated

“Each expert model is developed to become a “specialist” in a specific geography and use case”article, stated

“Only a high-level encoding of the fraud patterns, represented in the form of machine learning models, is shared across institutions.”article, stated

Why this architecture

The design targets two limits of conventional fraud tooling: rules and institution-specific models are slow to build and narrow to one institution's history, while raw cross-bank data sharing would violate privacy. MoE plus federated learning lets Feedzai reuse network intelligence without moving customer data, then calibrate scores so institutions can operationalize the risk signal in existing systems.

3 supporting quotes

“rule-based systems and custom AI models require substantial time and effort to design, tune, and validate before they can be deployed.”article, stated

“conventional detection systems are developed using knowledge from a single institution, offering a narrow and siloed view of fraud.”article, stated

“data never leaves the client’s secure environment. Only a high-level encoding of the fraud patterns”article, stated

Models used: Feedzai TrustScore · pre-trained fraud expert models · Mixture of Experts aggregation models

Scale

Feedzai describes the network as $8.02T in yearly processed events.
“our network of $8.02T yearly processed events”article, stated
Fraud can represent less than 0.1% of financial transactions.
“Fraud accounts for a tiny fraction of financial transactions, often less than 0.1%.”article, stated
Financial fraud/scam victims are described as losing over $2,000 on average.
“revealing an average loss of over $2,000 per victim. To protect their customers, banks and other financial institutions have made great efforts”article, stated
Traditional rules/custom AI may require months of analyst, data-science, and domain-expert work.
“This often involves months of work by analysts, data scientists, and domain experts.”article, stated

Alternatives considered (roads not taken)

Handcrafted rule-based fraud systems

Rejected because: Rules are rigid, slow to tune, and struggle to keep up with evolving fraud patterns.

“many financial institutions still rely on detection systems that are either too rigid (e.g., relying on handcrafted rules) or too narrow”article, stated

Institution-specific custom AI only

Rejected because: Single-institution models are reactive and blind to fraud patterns not yet seen locally.

“developed using knowledge from a single institution, offering a narrow and siloed view of fraud.”article, stated

“blind to emerging fraud patterns or schemes that haven’t yet appeared in their own environment.”article, stated

Share raw transaction data across institutions

Rejected because: TrustScore shares high-level fraud-pattern encodings rather than moving sensitive customer data.

“data never leaves the client’s secure environment. Only a high-level encoding of the fraud patterns”article, stated

Replace every existing fraud system

Rejected because: TrustScore can augment existing rules and custom models rather than requiring a rip-and-replace migration.

“this risk score can be seamlessly combined with other solutions, such as rules or custom AI models, to augment their performance.”article, stated

Patterns

`model-routing` (mixture-of-experts-risk-scoring)

TrustScore assembles relevant pre-trained expert models for a client and combines them into a final fraud-risk decision.

“Each expert model is developed to become a “specialist” in a specific geography and use case (e.g., banking, payment processing, or anti-money laundering).”article, stated

`privacy-preserving-processing` (federated-fraud-intelligence)

Raw transaction data remains inside each client's environment while trained model encodings share fraud-pattern knowledge.

“Only a high-level encoding of the fraud patterns, represented in the form of machine learning models, is shared across institutions.”article, stated

`data-quality-loop` (continuous-fraud-refresh)

TrustScore has built-in continuous refreshing so the system can adapt to new fraud patterns after deployment.

“this solution is ready-to-use, doesn’t require months of historical data collection, has built-in processes for continuous refreshing, and can leverage shared intelligence”article, stated

`observability-tracing` (expert-and-ensemble-monitoring)

The production system monitors individual expert models and the aggregate MoE system for drift and performance drops.

“The production system implements monitoring at two distinct levels: individual expert model performance and the aggregated MoE system performance.”summary, stated

`confidence-thresholding` (stable-risk-score-calibration)

Calibration keeps expert outputs and final scores consistent so downstream systems interpret risk levels reliably.

“We use calibration techniques to ensure that experts yield scores that represent fraud risk homogeneously.”article, stated

Tradeoffs

How to shorten time to fraud-detection value for new clients.

Chose pre-trained experts that work as a zero-day solution over waiting months for institution-specific historical data and custom model setup.

Rationale: The shared expert network provides immediate baseline detection while allowing later client-specific optimization.

“an AI-powered fraud risk score designed to detect fraud as a zero-day solution, without relying on historical data or complex model setups.”article, stated

How to introduce TrustScore into existing fraud stacks.

Chose augment rules and custom AI models over requiring a rip-and-replace deployment.

Rationale: The risk score can become an additional signal while institutions keep existing workflows.

“this risk score can be seamlessly combined with other solutions, such as rules or custom AI models, to augment their performance.”article, stated

Gotchas (what bit them)

Cross-institution fraud intelligence only works if data is standardized first.

Root cause: Institutions encode schemas, categories, currencies, and time zones differently.

Resolution: Feedzai standardizes schema, categorical values, currencies, and time-derived features before applying shared expert models.

“the format and the meaning of the data must be standardized, such that the patterns that the experts learned can be seamlessly integrated into new financial institutions.”article, stated

Independent expert models can distort the final ensemble if their scores are not calibrated.

Root cause: Each expert is trained in its own environment without context from the other experts.

Resolution: Calibration is used so every expert contributes a homogeneous fraud-risk representation.

“each expert is trained independently, without context from the other experts and their environments.”article, stated

Changing expert models can make downstream fraud scores unstable.

Root cause: Fraud patterns evolve and individual experts are updated, but operational systems expect consistent score meaning.

Resolution: Feedzai keeps score distributions stable as the underlying expert system changes.

“As fraud patterns evolve and expert models are updated, the risk score must remain consistent and reliable over time.”article, stated

Evaluation

Feedzai evaluates TrustScore through fraud detection lift, false-alert reduction, score calibration, score stability, and ongoing monitoring of individual experts and the aggregate model.

Metrics: fraud detection improvement · false alert reduction · risk score stability · expert calibration · time to market · model drift

“clients across geographies and use cases have experienced a significant performance improvement compared to traditional systems, as well as a reduction of time spent chasing false positives.”article, stated

“To ensure this, we apply techniques that keep the score distribution stable, even as the underlying system changes.”article, stated

Outcomes

TrustScore provides a zero-day fraud risk score using shared fraud intelligence.
Feedzai reports improved fraud detection, reduced false alerts, and faster time to market compared with traditional systems.
The system draws on a network processing $8.02T yearly events.

The article is a vendor-authored case and does not disclose exact client metrics, but it describes production deployment across geographies and financial use cases.

“our network of $8.02T yearly processed events”article, stated

“Real-world results show a boost in fraud detection, a reduction in false alerts, and a faster time to market”article, stated