Privacy-Preserving Detection of Ghost Job Listings on Freelance Platforms Using Federated Autoencoders

Abstract

The growing reliance on freelance platforms has led to a surge in ghost job listings—fake or misleading posts that waste freelancers' time, risk data privacy, and disrupt platform trust. Traditional centralized anomaly detection systems pose privacy concerns and struggle to generalize across diverse job markets. This paper proposes a novel federated learning framework that detects ghost job listings using local client-side training and privacy-preserving global model aggregation. Our approach integrates personality-based anomaly signals derived from job description text, capturing unusual tone, emotional inconsistency, and behavioral patterns typical of fraudulent listings. A federated autoencoder model, paired with summary statistics-based thresholding, enables robust anomaly detection across heterogeneous and non-IID client data. Experimental evaluations on synthetic and real-world datasets demonstrate that the proposed system outperforms traditional centralized models in both precision and recall while preserving user privacy.