How malicious servers exploit PEFT adapters to steal your training data
A new attack lets attackers reconstruct up to 79% of original training samples from seemingly safe PEFT adapters in federated learning. A lightweight tool now detects these backdoors before deployment.