Parallel Unlearning in Inherited Model Networks

TL;DR

This work addresses the challenge of unlearning in inherited model networks by modeling model relationships as a chronologically ordered DAG and introducing FIUn, a Fisher Information-based method for parallel unlearning. FIUn identifies critical last-layer parameters through FIM comparisons between unlearning and training data, enabling independent, parallel updates across inherited models, while MFIM merges multiple upstream FIMs to handle multi-root scenarios. Empirical results across CIFAR100, TinyImageNet, and Yahoo! Answers demonstrate near-zero forgetting for unlearned labels, strong retention for retained labels, and substantial speedups (up to 99%) compared with retraining or other baselines, with robust performance across varying inheritance depths and label distributions. The approach is applicable to FL, DDPL, IL, and TL settings, offering scalable, auditable, and efficient unlearning in complex model networks.

Abstract

Unlearning is challenging in generic learning frameworks with the continuous growth and updates of models exhibiting complex inheritance relationships. This paper presents a novel unlearning framework that enables fully parallel unlearning among models exhibiting inheritance. We use a chronologically Directed Acyclic Graph (DAG) to capture various unlearning scenarios occurring in model inheritance networks. Central to our framework is the Fisher Inheritance Unlearning (FIUn) method, designed to enable efficient parallel unlearning within the DAG. FIUn utilizes the Fisher Information Matrix (FIM) to assess the significance of model parameters for unlearning tasks and adjusts them accordingly. To handle multiple unlearning requests simultaneously, we propose the Merging-FIM (MFIM) function, which consolidates FIMs from multiple upstream models into a unified matrix. This design supports all unlearning scenarios captured by the DAG, enabling one-shot removal of inherited knowledge while significantly reducing computational overhead. Experiments confirm the effectiveness of our unlearning framework. For single-class tasks, it achieves complete unlearning with 0% accuracy for unlearned labels while maintaining 94.53% accuracy for retained labels. For multi-class tasks, the accuracy is 1.07% for unlearned labels and 84.77% for retained labels. Our framework accelerates unlearning by 99% compared to alternative methods. Code is in https://github.com/MJLee00/Parallel-Unlearning-in-Inherited-Model-Networks.

Figures (13)

• Figure 1: These diagrams show learning frameworks modeled as a DAG, where one user may create multiple models, and multiple models may be created by different users.
• Figure 2: The dark blue nodes are all model nodes inheriting from the starting nodes $n_{s}$ and $n_{f}$ in light blue. A model node $n_{\ast}$ corresponds to a model $w_{\ast}$. $w_{s}$ and $w_{f}$ can originate from different users, with the possibility of simultaneous unlearning requests.
• Figure 3: The unlearning process in $\mathcal{G}$ consists of five steps: 1) Preparation. The discovery nodes $\{n^{ D}_i\}$ are identified, with unlearning data $\hat{D}^{ D}_i$ containing one or more labels. Subsequently, the unlearning graph $\mathcal{\hat{G}}$ is located which root at $\{n^{ D}_i\}$. 2) Calculate unlearning the unlearning FIMs $\{\hat{F}_{i}^{ D}\}$ of $\{n^{ D}_i\}$. 3) Calculate the model FIM $F_{j}$ of node $n_j$ in $\mathcal{\hat{G}}$. 4) Merge unlearning FIMs $\hat{F}^{ D}_{j,k}$ for $n_j$ in the case of multiple discovery nodes according to the topology. 5) Update $n_j$ by comparing the merged unlearning FIM $\hat{F}_{j}^{ M}$ and $F_{j}$. Steps 2, 3, 4 and 5 apply to all nodes in $\mathcal{\hat{G}}$ .
• Figure 4: Three discovery nodes need to undergo the unlearning process, namely model $w_j$, model $w_s$ and model $w_f$. The labels to be unlearned are label 1 and label 2. However, model $w_j$ and model $w_s$ only have label 1, while model $w_f$ only has label 2. Model $w_h$ inherits from these three discovery nodes, so we merge the unlearning FIMs of the discovery nodes and perform the unlearning of label 1 and label 2 on model $w_h$ together.
• Figure 5: Models topologies considered in the experiments.