Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy Adhering Machine Un-learning in NLP

Vinayshekhar Bannihatti Kumar, Rashmi Gangadharaiah, Dan Roth

TL;DR

This work tackles the high cost of data deletion under privacy regulations by adapting the Sharded, Isolated, Sliced and Aggregated (SISA) unlearning framework to NLP. It introduces two NLP-specific variants, SISA-FC and SISA-A, to avoid storing full model checkpoints and dramatically reduce memory, time, and storage while preserving performance on GLUE tasks. Experimentally, SISA-A achieves near-BERT-base accuracy with about a 100x speedup in re-training and substantial memory savings, while SISA-FC offers even lower memory use at a smaller accuracy trade-off. The results demonstrate a practical path for privacy-compliant NLP systems, though current approaches rely on an aggregation layer that limits applicability to decoder-based architectures; future work will extend unlearning to those models and refine data-partitioning strategies.

Abstract

Regulations introduced by General Data Protection Regulation (GDPR) in the EU or California Consumer Privacy Act (CCPA) in the US have included provisions on the \textit{right to be forgotten} that mandates industry applications to remove data related to an individual from their systems. In several real world industry applications that use Machine Learning to build models on user data, such mandates require significant effort both in terms of data cleansing as well as model retraining while ensuring the models do not deteriorate in prediction quality due to removal of data. As a result, continuous removal of data and model retraining steps do not scale if these applications receive such requests at a very high frequency. Recently, a few researchers proposed the idea of \textit{Machine Unlearning} to tackle this challenge. Despite the significant importance of this task, the area of Machine Unlearning is under-explored in Natural Language Processing (NLP) tasks. In this paper, we explore the Unlearning framework on various GLUE tasks \cite{Wang:18}, such as, QQP, SST and MNLI. We propose computationally efficient approaches (SISA-FC and SISA-A) to perform \textit{guaranteed} Unlearning that provides significant reduction in terms of both memory (90-95\%), time (100x) and space consumption (99\%) in comparison to the baselines while keeping model performance constant.

Privacy Adhering Machine Un-learning in NLP

TL;DR

This work tackles the high cost of data deletion under privacy regulations by adapting the Sharded, Isolated, Sliced and Aggregated (SISA) unlearning framework to NLP. It introduces two NLP-specific variants, SISA-FC and SISA-A, to avoid storing full model checkpoints and dramatically reduce memory, time, and storage while preserving performance on GLUE tasks. Experimentally, SISA-A achieves near-BERT-base accuracy with about a 100x speedup in re-training and substantial memory savings, while SISA-FC offers even lower memory use at a smaller accuracy trade-off. The results demonstrate a practical path for privacy-compliant NLP systems, though current approaches rely on an aggregation layer that limits applicability to decoder-based architectures; future work will extend unlearning to those models and refine data-partitioning strategies.

Abstract

Regulations introduced by General Data Protection Regulation (GDPR) in the EU or California Consumer Privacy Act (CCPA) in the US have included provisions on the \textit{right to be forgotten} that mandates industry applications to remove data related to an individual from their systems. In several real world industry applications that use Machine Learning to build models on user data, such mandates require significant effort both in terms of data cleansing as well as model retraining while ensuring the models do not deteriorate in prediction quality due to removal of data. As a result, continuous removal of data and model retraining steps do not scale if these applications receive such requests at a very high frequency. Recently, a few researchers proposed the idea of \textit{Machine Unlearning} to tackle this challenge. Despite the significant importance of this task, the area of Machine Unlearning is under-explored in Natural Language Processing (NLP) tasks. In this paper, we explore the Unlearning framework on various GLUE tasks \cite{Wang:18}, such as, QQP, SST and MNLI. We propose computationally efficient approaches (SISA-FC and SISA-A) to perform \textit{guaranteed} Unlearning that provides significant reduction in terms of both memory (90-95\%), time (100x) and space consumption (99\%) in comparison to the baselines while keeping model performance constant.
Paper Structure (19 sections, 1 equation, 13 figures, 1 table)

This paper contains 19 sections, 1 equation, 13 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets
  4. Models
  5. SISA
  6. SISA modified with fully connected Layer (SISA-FC)
  7. SISA modified with Adapters (SISA-A)
  8. Baselines:
  9. Implementation details:
  10. Experiments and Results
  11. Evaluation Metrics:
  12. Analysis of SISA-A
  13. Analysis of SISA-FC
  14. Comparison Between SISA-FC and SISA-A
  15. Profiling requests based on probability of occurrence
  16. ...and 4 more sections

Figures (13)

  • Figure 1: Architecture of SISA-A which uses S slices and R shards. One model is built for each shard and the labels are aggregated using a majority voting strategy.
  • Figure 5: Performance of SISA-A on the 3 datasets.
  • Figure 9: Re-training time of SISA-A on the 3 datasets.
  • Figure 13: Performance of SISA-FC on the 3 datasets.
  • Figure 17: Re-training time of SISA-FC on the 3 datasets.
  • ...and 8 more figures