From beats to leaks: how offensive AI infers users' sensitive information from playlists

The researchers propose musicPIIrate, which can infer sensitive information such as age, country, gender, drinking and smoking and exercise habits, and OCEAN personality scores from users' public playlists; the authors claim it achieves current best inference accuracy. The method combines set models such as Deep Sets with graph neural networks to model data within playlists and relationships between playlists, and outperforms baselines on 9 of 15 attribute inference tasks. To mitigate the risk, the authors propose JamShield, which dilutes the PII-carrying signal by injecting fake playlists into accounts; experiments show inference F1 drops by 10% on average.