Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks

Jing Wu, Mehrtash Harandi

TL;DR

Scissorhands addresses the problem of erasing a forgetting data's influence from trained models under data-privacy regulations. It introduces a two-stage workflow that first trims data-influential parameters using connection sensitivity (top-$k$%) and then repairs the model with a gradient projection approach to maximize remaining-data utility while suppressing forgetting-data influence. The method demonstrates competitive forgetting-utility trade-offs across image classification and generation tasks, including a Stable Diffusion case study, supported by ablations and case analyses. This work provides a scalable, principled approach to privacy-preserving unlearning with practical implications for deployment and policy compliance.

Abstract

Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top-k percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands.

Scissorhands: Scrub Data Influence via Connection Sensitivity in Networks

TL;DR

Scissorhands addresses the problem of erasing a forgetting data's influence from trained models under data-privacy regulations. It introduces a two-stage workflow that first trims data-influential parameters using connection sensitivity (top-%) and then repairs the model with a gradient projection approach to maximize remaining-data utility while suppressing forgetting-data influence. The method demonstrates competitive forgetting-utility trade-offs across image classification and generation tasks, including a Stable Diffusion case study, supported by ablations and case analyses. This work provides a scalable, principled approach to privacy-preserving unlearning with practical implications for deployment and policy compliance.

Abstract

Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new machine unlearning approach Scissorhands. Initially, Scissorhands identifies the most pertinent parameters in the given model relative to the forgetting data via connection sensitivity. By reinitializing the most influential top-k percent of these parameters, a trimmed model for erasing the influence of the forgetting data is obtained. Subsequently, Scissorhands fine-tunes the trimmed model with a gradient projection-based approach, seeking parameters that preserve information on the remaining data while discarding information related to the forgetting data. Our experimental results, conducted across image classification and image generation tasks, demonstrate that Scissorhands, showcases competitive performance when compared to existing methods. Source code is available at https://github.com/JingWu321/Scissorhands.
Paper Structure (25 sections, 1 theorem, 12 equations, 8 figures, 9 tables, 1 algorithm)

This paper contains 25 sections, 1 theorem, 12 equations, 8 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Overview.
  4. Trimming
  5. Repairing
  6. Algorithm Description
  7. Related Work
  8. Experimental Evaluation
  9. Results on Classification Task
  10. Sample-wise unlearning.
  11. Class-wise unlearning.
  12. Ablation Study
  13. Component analysis.
  14. Hyper-parameter impact.
  15. Case Study: Stable Diffusion
  16. ...and 10 more sections

Key Result

lemma thmcounterlemma

If the primal problem has the optimal solution $f^\ast$ and its dual problem has the optimal solution $h^\ast$, then $h^\ast = \sup_{v} \inf_{{\bm{g}}} L({\bm{g}}, v) \leq \inf_{{\bm{g}}} \sup_{v} L({\bm{g}}, v) = f^\ast$.

Figures (8)

  • Figure 1: Visualizations of regions where models focus on generated by GradCAM selvaraju2017grad.
  • Figure 2: Influence of the percent value $k$ in the trimming process, the balance term $\lambda$ of \ref{['eq:obj']}, and the ratio of forgetting data used in the trimming process.
  • Figure 3: Quantity of nudity content detected using the NudeNet classifier from 1K sampled images and I2P data. We observed a high false positive rate for exposed female genitalia/breast using the NudeNet classifier on generated I2P images. The flagged images can be found in Appendix B.
  • Figure 4: Sample images with the prompt from $c_f=${'nudity', 'naked', 'erotic', 'sexual'} generated by SDs w/ and w/o machine unlearning algorithms. Best viewed in color.
  • Figure 5: Visualizations of regions where models focus on generated by GradCAM selvaraju2017grad. Best viewed in color.
  • ...and 3 more figures

Theorems & Definitions (7)

  • remark thmcounterremark
  • remark thmcounterremark
  • remark thmcounterremark
  • remark thmcounterremark
  • proof
  • lemma thmcounterlemma
  • proof