Multi-Label Red Flag Detection System for Fraud Prevention
feedzai · Finance · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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: Fraud and scam detection
Feedzai needed fraud detection that explains observable scam red flags instead of returning only opaque binary scam scores.
Constraints:
- A request can appear suspicious in isolation while being legitimate with user context.
- Users need interpretable warnings rather than absolute AI judgments.
- The system must handle screenshots from emails, text messages, listings, and other communications.
- Commercial multimodal models vary in accuracy, latency, cost, and instruction following.
Architecture
ScamAlert uses multimodal LLMs to analyze screenshots, detect multiple red flags from a predefined taxonomy, explain why each flag applies, recommend next steps, and benchmark model choices across accuracy, instruction adherence, cost, and latency.
Diagram confidence: reconstructed (reconstructed from prose).
flowchart LR screenshot-input["Screenshot Input"] multimodal-llm["Multimodal LLM"] red-flag-taxonomy["Red Flag Taxonomy"] structured-output-layer["Structured Output Layer"] benchmarking-framework["Benchmarking Framework"] screenshot-input --> multimodal-llm multimodal-llm --> red-flag-taxonomy multimodal-llm --> structured-output-layer red-flag-taxonomy --> multimodal-llm red-flag-taxonomy --> benchmarking-framework structured-output-layer --> multimodal-llm benchmarking-framework --> multimodal-llm benchmarking-framework --> red-flag-taxonomy
Components
| Component | Role | Interacts with |
|---|---|---|
Screenshot Input |
Captures suspected scam communications as images. | Multimodal LLM |
Multimodal LLM |
Analyzes screenshots and predicts red flags from the taxonomy. | Red Flag Taxonomy, Structured Output Layer |
Red Flag Taxonomy |
Defines allowed labels such as suspicious links, urgency tactics, and unusual channels. | Multimodal LLM, Benchmarking Framework |
Structured Output Layer |
Returns red flags, explanations, and next-step recommendations. | Multimodal LLM |
Benchmarking Framework |
Evaluates models across accuracy, instruction adherence, cost, and latency. | Multimodal LLM, Red Flag Taxonomy |
Data flow
A user submits a screenshot; the multimodal model identifies taxonomy-constrained red flags, produces explanations and recommendations, and the benchmark pipeline evaluates detection quality, parsing validity, hallucinated flags, cost, and latency.
Why this architecture
Multi-label red-flag detection gives users concrete, inspectable reasons while avoiding overconfident binary judgments in ambiguous fraud contexts.
Models used: GPT-5 · Gemini 3 Pro · Gemini 2.5 Pro · commercial multimodal models
Scale
- The benchmark measures four production dimensions.“evaluate multiple commercial multimodal models across four dimensions: red flag detection accuracy (precision/recall/F1), instruction adherence, cost, and latency.”summary, stated
- Each screenshot was evaluated three times to capture generative variability.“The team ran evaluations three times per screenshot to account for the inherent variability in generative models”summary, stated
Alternatives considered (roads not taken)
Traditional binary scam classification
Rejected because: Binary classifiers output risk scores without interpretable user-facing reasons.
Free-form generated warning labels
Rejected because: Models can hallucinate invalid red flags unless constrained to a predefined taxonomy.
Patterns
eval-driven-development (quality)
Feedzai built a benchmark spanning recall, precision, F1, instruction adherence, cost, and latency.
safety-policy-enforcement (governance)
The red-flag taxonomy constrains model outputs and prevents hallucinated labels from being accepted.
human-in-the-loop-gate (governance)
The system assists user awareness instead of making absolute decisions, leaving context-aware judgment to users.
Tradeoffs
Use multi-label red flags
Chose Multiple observable taxonomy labels with explanations over Single binary scam score.
Rationale: This improves transparency and supports ambiguous fraud contexts.
Benchmark beyond accuracy
Chose Accuracy, instruction adherence, cost, and latency over Choosing the most accurate model only.
Rationale: Production deployment depends on behavior, parseability, price, and response time.
Gotchas (what bit them)
Instruction-following failures can look like detection success unless taxonomy validity is checked.
Root cause: Some models generate red flags outside the predefined taxonomy.
Resolution: Treat formatting errors and invalid predictions as benchmark failures.
Binary fraud judgments can be misleading without user context.
Root cause: The user may know context that the classifier cannot see.
Resolution: Return observable warning signs and explanations instead of absolute scam determinations.
Evaluation
Feedzai manually annotated screenshots, evaluated recall/precision/F1, instruction adherence, cost, latency, and repeated runs for variability.
Metrics: precision · recall · F1 · instruction adherence · cost per query · latency
Outcomes
- Feedzai built ScamAlert as an explainable multimodal red-flag detector.
- The benchmark found large model-to-model differences in accuracy, cost, latency, and instruction adherence.
- GPT-5, Gemini 3 Pro, and Gemini 2.5 Pro led reported accuracy.
The case emphasizes benchmark dimensions and repeated screenshot evaluation rather than live traffic volume.