safetyfactbullishA model-agnostic framework can jointly address privacy preservation and malicious behavior across both federated and decentralized learning settingsMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02187v1