Key protected classification for collaborative learning

Mert Bülent Sariyildiz, Ramazan Gokberk Cinbis, Erman Ayday

Abstract: Highlights•Collaborative learning allow privacy-preserving training in sensitive domains.•Collaborative learning is vulnerable to generative adversarial training-based attacks.•An attack-resilient classification model and principled training scheme is proposed.•The proposed model prevents active attacks by hiding class scores via class keys.•How to utilize high-dimensional keys without increasing model complexity is shown.