PUMA: Performance Unchanged Model Augmentation for Training Data Removal

Ga Wu, Masoud Hashemi, Christopher Srinivasa

Abstract

Preserving the performance of a trained model while removing unique characteristics of marked training data points is challenging. Recent research usually suggests retraining a model from scratch with remaining training data or refining the model by reverting the model optimization on the marked data points. Unfortunately, aside from their computational inefficiency, those approaches inevitably hurt the resulting model's generalization ability since they remove not only unique characteristics but also discard shared (and possibly contributive) information. To address the performance degradation problem, this paper presents a novel approach called Performance Unchanged Model Augmentation (PUMA). The proposed PUMA framework explicitly models the influence of each training data point on the model's generalization ability with respect to various performance criteria. It then complements the negative impact of removing marked data by reweighting the remaining data optimally. To demonstrate the effectiveness of the PUMA framework, we compared it with multiple state-of-the-art data removal techniques in the experiments, where we show the PUMA can effectively and efficiently remove the unique characteristics of marked training data without retraining the model that can 1) fool a membership attack, and 2) resist performance degradation. In addition, as PUMA estimates the data importance during its operation, we show it could serve to debug mislabelled data points more efficiently than existing approaches.

Introduction

As data privacy regulations like GDPR and CCPA grant individuals the "right to be forgotten," machine learning researchers face a daunting technical challenge: Machine Unlearning. The objective is to remove the influence of specific training data points from a model without the prohibitive computational cost of retraining from scratch.

Existing "fast" removal methods often fall into two traps. First, they frequently cause performance degradation by removing shared information that aids generalization along with the unique characteristics of the deleted data. Second, methods like SISA (Sharded, Isolated, Sliced, and Aggregated training) introduce significant logistical complexity, requiring the maintenance of numerous sub-models. PUMA, introduced by Ga Wu and colleagues from Borealis AI, addresses these issues by "patching" an existing model to forget specific data while ensuring its overall performance remains stable.

Methodology: Optimal Reweighting

PUMA approaches data removal as a constrained optimization problem. Its methodology relies on two primary pillars:

  1. Influence Functions: The framework uses influence functions to estimate how much each individual training point contributed to the model's final parameters and its performance.
  2. Optimal Reweighting (The "Patch"): Instead of simply subtracting the influence of the "marked" data—which typically leads to a performance drop—PUMA re-evaluates the remaining data. It calculates a new set of weights for the remaining points, optimized to "fill the hole" left by the removed data and compensate for lost performance.
Comparison of removal techniques: (a) original model, (b) naive removal showing performance drift, and (c) PUMA maintaining performance through reweighting.

Technically, this involves a linear optimization task using Hessian Vector Products (HVP) to maintain computational efficiency. The final step is a simple "patching" operation that updates the model parameters linearly.

Experimental Evaluation

The authors evaluated PUMA across synthetic data, tabular datasets (UCI German Credit, Breast Cancer), and image data (MNIST).

Initial model prediction on synthetic radial cluster data before removal.

Key Results

  • Superior Accuracy: In tests removing up to 80% of data, PUMA consistently maintained higher accuracy than baselines like Amnesiac ML or SISA.
  • Privacy Protection: PUMA effectively resisted Membership Inference Attacks, reducing the identification rate of removed data to near-zero or baseline levels.
  • Efficiency: PUMA is significantly faster than retraining and dominates SISA in "random" removal scenarios—the most common real-world use case.
  • State-of-the-Art Debugging: PUMA outperformed established methods like Data Shapley in identifying mislabeled data points while inspecting a smaller fraction of the dataset.

Conclusion and Future Work

PUMA represents a significant advancement in machine unlearning, demonstrating that linear approximations can effectively "patch" models for data removal. While the method is constrained by the limits of first-order Taylor expansions and requires careful hyperparameter tuning, it provides a scalable and performance-preserving alternative to retraining. Future work remains to explore its application to modern Large Language Models with billions of parameters.

Summary updated

Abstract

Preserving the performance of a trained model while removing unique characteristics of marked training data points is challenging. Recent research usually suggests retraining a model from scratch with remaining training data or refining the model by reverting the model optimization on the marked data points. Unfortunately, aside from their computational inefficiency, those approaches inevitably hurt the resulting model's generalization ability since they remove not only unique characteristics but also discard shared (and possibly contributive) information. To address the performance degradation problem, this paper presents a novel approach called Performance Unchanged Model Augmentation (PUMA). The proposed PUMA framework explicitly models the influence of each training data point on the model's generalization ability with respect to various performance criteria. It then complements the negative impact of removing marked data by reweighting the remaining data optimally. To demonstrate the effectiveness of the PUMA framework, we compared it with multiple state-of-the-art data removal techniques in the experiments, where we show the PUMA can effectively and efficiently remove the unique characteristics of marked training data without retraining the model that can 1) fool a membership attack, and 2) resist performance degradation. In addition, as PUMA estimates the data importance during its operation, we show it could serve to debug mislabelled data points more efficiently than existing approaches.
BibTeX
@inproceedings{wu2022puma,
  author = {Ga Wu and Masoud Hashemi and Christopher Srinivasa},
  title = {{PUMA: Performance Unchanged Model Augmentation for Training Data Removal}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=nJha8GBTtn},
}
Figure