De-identification in game chargeback fraud detection

Online games suffer from chargeback (malicious refund) abuse. Training effective detectors often requires high-utility telemetry that contains quasi-identifiers (QIs), creating privacy risks. We present a risk-driven pipeline: quantify re-identification risk, apply ISO/IEC 20889 de-identification controls, then measure downstream utility for refund-fraud detection. We compare k-anonymity (generalization/suppression) and DP-based synthesis, and evaluate stream models (ARF, SAM-kNN) and sequence models (LSTM/GRU) with Matthews correlation coefficient (MCC). Moderate k (10–20) preserves strong utility for sequence models while materially lowering risk; DP synthesis offers stronger formal privacy but reduces utility the most.