In essence, MAE proposed an asymmetric encoder-decoder architecture for MIM, where the encoder (e.g., a standard ViT model [17])operates onlyonvisible patches, andthelight-weight decoder recoversallpatches for maskprediction.

Existing methods employ the influence function to achieve feature unlearning, which is impractical for FL as it necessitates the participation of other clients,if not all, in the unlearning process. Furthermore, current research lacks an evaluation of the effectiveness of feature unlearning. Toaddress these limitations, we define feature sensitivity in evaluating feature unlearning according to Lipschitz continuity. Thismetric characterizes themodel output'srateofchange or sensitivity to perturbations in the input feature. We then propose an effective federated feature unlearning framework called Ferrari, which minimizes feature sensitivity. Extensive experimental results and theoretical analysis demonstrate the effectiveness of Ferrari across various feature unlearning scenarios, including sensitive, backdoor, and biased features.