Datasets

Dataset Overview

This dataset supports our ASONAM 2025 study on the authenticity of LinkedIn profiles. It contains 4,200 profiles distributed as follows:

• 1,800 Legitimate LinkedIn Profiles (LLPs)

• 600 Manual Fake LinkedIn Profiles (FLPs)

• 1,200 GPT-3.5-Generated Profiles (GPT3.5Ps)

• 600 GPT-4-Generated Profiles (GPT4Ps)

It is the first dataset designed to test profile detection models against real, manually faked, and LLM-generated adversarial examples.

Profile Categories

1. Legitimate LinkedIn Profiles (LLPs)

• Count: 1,800

• Source: Public LinkedIn profiles, visible without login

• Includes: Name, Job Titles, Experience, Education, About, Skills, Followers

2. Manual Fake Profiles (FLPs)

• Count: 600

• Collected from: User reports, hashtag anomalies, fake org listings

• Features: Content reuse, inflated roles, buzzwords, cloned bios

3. GPT-3.5 Fake Profiles (GPT3.5Ps)

• Count: 1,200

• Generated using zero-shot prompting

• Sections include: Job Titles, Companies, Education, About

• Quality: Reasonable structure, less coherence than GPT-4

4. GPT-4 Fake Profiles (GPT4Ps)

• Count: 600

• Generated using few-shot prompting with real LLPs

• Enforced diversity: Region, industry, role, org type

• Includes all standard fields; highly realistic output

Data Structure and Features

Each profile includes both textual and numeric information:

• Sections: About, Experience, Education

• Subsections: Titles, companies, durations, institutions

• Numerical features: Experience count, skill count, followers, connections

All profiles are consistently structured and formatted for model ingestion or manual inspection.

Similarity Metrics and Preprocessing

Cosine Similarity

We measured textual similarity between generated and legitimate profiles: