Jellyfish: Zero-Shot Federated Unlearning Scheme with Knowledge Disentanglement

Abstract

With the increasing importance of data privacy and security, federated unlearning emerges as a new research field dedicated to ensuring that once specific data is deleted, federated learning models no longer retain or disclose related information. In this paper, we propose a zero-shot federated unlearning scheme, named Jellyfish. It distinguishes itself from conventional federated unlearning frameworks in four key aspects: synthetic data generation, knowledge disentanglement, loss function design, and model repair. To preserve the privacy of forgotten data, we design a zero-shot unlearning mechanism that generates error-minimization noise as proxy data for the data to be forgotten. To maintain model utility, we first propose a knowledge disentanglement mechanism that regularises the output of the final convolutional layer by restricting the number of activated channels for the data to be forgotten and encouraging activation sparsity. Next, we construct a comprehensive loss function that incorporates multiple components, including hard loss, confusion loss, distillation loss, model weight drift loss, gradient harmonization, and gradient masking, to effectively align the learning trajectories of the objectives of ``forgetting" and ``retaining". Finally, we propose a zero-shot repair mechanism that leverages proxy data to restore model accuracy within acceptable bounds without accessing users' local data. To evaluate the performance of the proposed zero-shot federated unlearning scheme, we conducted comprehensive experiments across diverse settings. The results validate the effectiveness and robustness of the scheme.

Figures (8)

• Figure 1: The motivation for knowledge disentanglement. The upper diagram illustrates the phenomenon of inter-class feature mixing during the feature extraction process in conventional models. This mixing leads to the simultaneous activation of features for cats and dogs in the activation maps (as shown in the red areas), affecting the accurate representation of single-class features. The lower diagram demonstrates the effect after applying knowledge disentanglement, where the model learns to represent features with less mixing. The activation maps show clearer class specificity (blue is primarily related to cats, and yellow to dogs). This disentanglement reduces activation overlap between different classes, enhancing the model's interpretability.
• Figure 2: Proposed Jellyfish Scheme: ① Noise Training: The user requests data deletion locally and trains proxy data $N_f$ to replace the deleted data. ② Knowledge Disentanglement: When the server receives the data, it first disentangles the data to reduce knowledge entanglement between categories. ③ Unlearn: The model is guided to forget by using an improved loss function and mechanisms like gradient harmonization. ④ Repair: If the model's accuracy drops after unlearning, the proxy data of the remaining data $N_r$ is used to restore the model's performance.
• Figure 3: Performance Metrics on Test Set. (a) JSD on Test Set, (b) L2 on Test Set, (c) p-value on Test Set.
• Figure 4: Comparison of accuracy on remaining and forgotten data using real and proxy data for unlearning. (a) accuracy on remaining data, and (b) accuracy on forgotten data.
• Figure 5: Accuracy of non-target categories before and after disentangling.