SCAMBENCH is a dataset for studying online scam communication, scam detection, and model robustness. It contains real-world scam messages curated from victim-reported incidents, along with non-scam counterparts and structured annotations.

The dataset is introduced in our paper:

SCAMBENCH: A Multi-Perspective Benchmark for Analyzing and Evaluating Online Scam Communication

Dataset Overview

SCAMBENCH consists of:

1,836 scam messages across six scam types
1,834 non-scam messages from legitimate sources
Multi-perspective annotations capturing:
- persuasion strategies (scammer perspective)
- red flags (user perspective)
- severity levels (risk and urgency)

In addition, each scam message includes a rewritten version where explicit scam cues are reduced, enabling robustness evaluation.

Dataset Structure

ScamBench/
├── Scam/
│   ├── Advance_fee_loan.csv
│   ├── Employment.csv
│   ├── Imposter.csv
│   ├── Tech_support.csv
│   ├── Lottery.csv
│   ├── Investment.csv
│
├── Non-scam/
│   ├── Employment.csv
│   ├── Imposter.csv
│   ├── Tech_support.csv
│   ├── Lottery.csv
│   ├── Investment.csv

Scam Data Format

Each CSV file in the Scam/ folder contains the following fields:

Field	Description
`description`	Original scam message
`rewrite_description`	Rewritten version with reduced persuasion/red-flag cues
`dollars_lost`	Reported financial loss (if available)
`location`	Reported location (may be anonymized or missing)
`date_reported`	Report date
`business_name_used`	Name used by scammer (if present)
`annotation`	JSON object with structured labels
`category`	Six categories (advance fee loan, employment, imposter, investment, lottery, and tech support)

Non-Scam Data Format

Each CSV file in the Non-scam/ folder contains the following fields:

Field	Description
`description`	Legitimate message
`category`	Five categories (employment, imposter, investment, lottery, and tech support)

Annotation Schema

The dataset provides three types of annotations:

1. Persuasion Strategies (P1–P7)

Based on Cialdini’s persuasion principles:

P1 Reciprocity – Creating obligation by offering help, rewards, gifts, or favors.
P2 Commitment & Consistency – Starting with a small request and escalating to larger demands.
P3 Social Proof – Claiming many others have benefited; using testimonials or group behavior.
P4 Authority – Invoking titles, institutions, official seals, or legal threats; impersonating authority.
P5 Liking – Using flattery, friendliness, shared interests, or emotional bonding.
P6 Scarcity – Using urgency, limited-time offers, countdowns, or “last chance” framing.
P7 Unity – Claiming shared identity (community, religion, nationality, ethnicity) to build trust.

2. Red Flags (R1–R9)

User-facing warning signals:

R1 Unsolicited contact / pitches
R2 “Too good to be true” / guaranteed outcomes
R3 Pressure to act now
R4 Requests personal or financial information
R5 Upfront fees / payments required
R6 Payment via gift card / wire / crypto
R7 Posing as government or official organization
R8 Threats of arrest / legal action
R9 Grammar / spelling / language issues

3. Severity Levels

Low
Medium
High
Critical

These annotations allow analysis beyond binary scam classification.

Use Cases

SCAMBENCH can be used for:

Scam detection and classification
Robustness evaluation (using rewritten messages)
Explainable AI and risk modeling
Human-centered security and scam education
Analysis of scam communication patterns

Ethical Considerations

The dataset is derived from publicly available scam reports.
Rewritten scam messages are generated for research purposes only, to evaluate detection robustness.
This dataset should not be used to generate or facilitate scams.

License

This dataset is released under the CC BY-NC 4.0 license.

Downloads last month: 66

Total file size:

10.2 MB