CallTrap Artifact

This repository accompanies the paper "When Scammers Talk Back: Understanding Potentially Unwanted Calls via LLM-Based Interaction" (CCS 2026).

Contents

• system_implementation.py: the sanitized system implementation used to operate CallTrap. Deployment-specific values remain as placeholders.
• coded_dataset/: the redacted coded dataset, divided into five JSON files. Together, the files contain 2,429 call records.
• codebook.pdf: the original artifact codebook.
• supplementary_material/full_appendix.pdf: the full appendix, including the classification criteria, system prompt, codebook, persuasion-technique definitions, method details, ethical discussion, and representative conversation examples.

System Implementation

system_implementation.py is provided as the sanitized research implementation. API keys, service credentials, public hostnames, routes, prompts, and voice identifiers are represented by placeholder values. The file is provided as-is and requires users to supply their own service configuration before deployment.

Redacted Coded Dataset

The five files in coded_dataset/ preserve the JSON organization used in the original artifact. Records are grouped by call-duration bucket. Caller names, organization names, phone numbers, and other sensitive information have been redacted; redacted values are represented as *****.

The camera-ready dataset uses the final coding framework. The persuasion field contains the original persuasion categories (auth, crc, dis, lsd, and sp). The complementary_persuasion field contains the added categories:

• nag: Need and Greed (NG in the paper)
• time: Time
• dh: Dishonesty (DSH in the paper)

For compatibility with the original framework, retained nag and time annotations are also mapped to dis, while retained dh annotations are mapped to crc in the original persuasion field.

Supplementary Material

The full appendix contains:

1. Open-science and ethical considerations
2. Generative-AI usage disclosure
3. PUC call-classification criteria
4. The complete system prompt with synthetic placeholders
5. The annotation codebook
6. Persuasion-technique definitions
7. Call topics and information-sensitivity tiers
8. The sampling-distribution comparison
9. The campaign-identification algorithm
10. Representative conversation examples