Federated learning (FL) enables collaborative model training with local data privacy preserving, but is vulnerable to backdoor attacks from malicious clients. These attacks can manipulate the global model to produce malicious output when encountering specific triggers. Current research on backdoor defense can be broadly classified into two categories based on the timing of defense: backdoor detection before aggregation and backdoor purification after aggregation. However, these approaches have limitations such as reliance on impractical assumptions like auxiliary data availability, susceptibility to inference attacks, and instability under non-independent and identically distributed (Non-IID) data. Therefore, there is a growing need for a backdoor-robust FL framework that can effectively suppress backdoor behavior while preserving task performance, even when a large fraction of clients launch backdoor attacks that closely mimic benign behavior. The defender must work without inspecting raw client data, without any auxiliary dataset, and without knowledge of attacker identities or trigger patterns, yet still adhere to strict privacy protocols.