Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LEGATO: Good Identity Unlearning Is Continuous

Qiang Chen, Chun-Wun Cheng, Xiu Su, Hongyan Xu, Xi Lin, Shan You, Angelica I. Aviles-Rivero, Yi Chen

TL;DR

LEGATO tackles the problem of forgetting a target identity in pretrained generative models by reframing unlearning as a continuous latent-space trajectory. It introduces lightweight Neural ODE adapters that modify only a small set of parameters while freezing the base generator, with forgetting strength controlled by the ODE step size and stabilized by a trajectory-consistency loss. The approach provides theoretical guarantees for smooth trajectories, non-intersecting identity flows, and conflict-free multi-identity unlearning, and demonstrates state-of-the-art forgetting with substantial parameter efficiency and faster updates in both in-domain and out-of-domain settings. This work advances privacy-preserving machine learning for large generative models by offering a scalable, interpretable, and robust unlearning framework with practical impact on rights-to-forget and data governance.

Abstract

Machine unlearning has become a crucial role in enabling generative models trained on large datasets to remove sensitive, private, or copyright-protected data. However, existing machine unlearning methods face three challenges in learning to forget identity of generative models: 1) inefficient, where identity erasure requires fine-tuning all the model's parameters; 2) limited controllability, where forgetting intensity cannot be controlled and explainability is lacking; 3) catastrophic collapse, where the model's retention capability undergoes drastic degradation as forgetting progresses. Forgetting has typically been handled through discrete and unstable updates, often requiring full-model fine-tuning and leading to catastrophic collapse. In this work, we argue that identity forgetting should be modeled as a continuous trajectory, and introduce LEGATO - Learn to ForgEt Identity in GenerAtive Models via Trajectory-consistent Neural Ordinary Differential Equations. LEGATO augments pre-trained generators with fine-tunable lightweight Neural ODE adapters, enabling smooth, controllable forgetting while keeping the original model weights frozen. This formulation allows forgetting intensity to be precisely modulated via ODE step size, offering interpretability and robustness. To further ensure stability, we introduce trajectory consistency constraints that explicitly prevent catastrophic collapse during unlearning. Extensive experiments across in-domain and out-of-domain identity unlearning benchmarks show that LEGATO achieves state-of-the-art forgetting performance, avoids catastrophic collapse and reduces fine-tuned parameters.

LEGATO: Good Identity Unlearning Is Continuous

TL;DR

LEGATO tackles the problem of forgetting a target identity in pretrained generative models by reframing unlearning as a continuous latent-space trajectory. It introduces lightweight Neural ODE adapters that modify only a small set of parameters while freezing the base generator, with forgetting strength controlled by the ODE step size and stabilized by a trajectory-consistency loss. The approach provides theoretical guarantees for smooth trajectories, non-intersecting identity flows, and conflict-free multi-identity unlearning, and demonstrates state-of-the-art forgetting with substantial parameter efficiency and faster updates in both in-domain and out-of-domain settings. This work advances privacy-preserving machine learning for large generative models by offering a scalable, interpretable, and robust unlearning framework with practical impact on rights-to-forget and data governance.

Abstract

Machine unlearning has become a crucial role in enabling generative models trained on large datasets to remove sensitive, private, or copyright-protected data. However, existing machine unlearning methods face three challenges in learning to forget identity of generative models: 1) inefficient, where identity erasure requires fine-tuning all the model's parameters; 2) limited controllability, where forgetting intensity cannot be controlled and explainability is lacking; 3) catastrophic collapse, where the model's retention capability undergoes drastic degradation as forgetting progresses. Forgetting has typically been handled through discrete and unstable updates, often requiring full-model fine-tuning and leading to catastrophic collapse. In this work, we argue that identity forgetting should be modeled as a continuous trajectory, and introduce LEGATO - Learn to ForgEt Identity in GenerAtive Models via Trajectory-consistent Neural Ordinary Differential Equations. LEGATO augments pre-trained generators with fine-tunable lightweight Neural ODE adapters, enabling smooth, controllable forgetting while keeping the original model weights frozen. This formulation allows forgetting intensity to be precisely modulated via ODE step size, offering interpretability and robustness. To further ensure stability, we introduce trajectory consistency constraints that explicitly prevent catastrophic collapse during unlearning. Extensive experiments across in-domain and out-of-domain identity unlearning benchmarks show that LEGATO achieves state-of-the-art forgetting performance, avoids catastrophic collapse and reduces fine-tuned parameters.
Paper Structure (41 sections, 6 theorems, 33 equations, 8 figures, 21 tables)

This paper contains 41 sections, 6 theorems, 33 equations, 8 figures, 21 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Formulation
  5. Method Overview
  6. Neural ODE Adapter for Unlearning
  7. Trajectory Consistency Constraint
  8. Conflict-Free Multi-Identity Unlearning
  9. Experimental Results
  10. Experimental Setting
  11. Overall Results
  12. Ablation and Sensitivity Studies
  13. Conclusion
  14. Acknowledgments
  15. Proof of Theorem 1
  16. ...and 26 more sections

Key Result

Theorem 1

Let $\Phi_{t_0 \rightarrow t} : [t_0, T) \times \mathbb{R}^d \times \Theta \rightarrow \mathbb{R}^d$, defined by $\Phi_{t_0 \rightarrow t}(x_0, \theta) = \varphi(t; t_0, x_0, \theta)$, be the solution map of a Neural ODE parameterized by $\theta$. If $f$ is Lipschitz continuous in $x$ and continuous

Figures (8)

  • Figure 1: An overview of LEGATO. LEGATO introduces fine-tuned Neural ODE with fewer parameters, instead of fine-tuning the pretrained generator. Stable forgetting is achieved by imposing trajectory consistency constraint on the function. LEGATO aims to push the identity of the forget set toward a different one while preserving the generative ability for retained identities.
  • Figure 1: Qualitative results of LEGATO in generative identity unlearning task. For each identity in the CelebAHQ dataset, the first row shows the source image and other images of the same identity, and the second row displays the results after forgetting the specific identity. The identities are sequentially 1784, 3478, 7901 and 55.
  • Figure 2: Qualitative results of GUIDE and the baseline in generative identity unlearning task. For the given source image each (the first row), LEGATO aimed to erase the identity in the pre-trained generator while preserving the ability to generate other identities. The images in the second and third row are the target and unlearned images, respectively.
  • Figure 2: Qualitative results of LEGATO and the baseline on a multi-image test using CelebAHQ dataset.
  • Figure 3: The Impact of TC on the retention loss.
  • ...and 3 more figures

Theorems & Definitions (10)

  • Theorem 1: Smooth Neural ODE Trajectories
  • Theorem 2: Non-Monotonicity of Step Size
  • Theorem 3: Conflict-Free Multi-Identity Unlearning
  • Remark 1
  • Theorem 1: Smooth Neural ODE Trajectories
  • proof
  • Theorem 2: Non-Monotonicity of Step Size
  • proof
  • Theorem 3: Conflict-free Multi-Identity Unlearning
  • proof