FedDetox: Robust Federated SLM Alignment via On-Device Data Sanitization

arXiv·medium signal

FedDetox tackles a critical problem in federated learning for small language models: real-world client data often contains toxic or unsafe content that poisons alignment. The framework performs on-device data sanitization before federated aggregation, filtering toxic samples without exposing raw data to the server. As federated fine-tuning becomes the primary path for private data utilization with scarce public data, this addresses a concrete deployment blocker.