Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Model Sparsity Can Simplify Machine Unlearning (2304.04934v13)

Published 11 Apr 2023 in cs.LG

Abstract: In response to recent data regulation requirements, machine unlearning (MU) has emerged as a critical process to remove the influence of specific examples from a given model. Although exact unlearning can be achieved through complete model retraining using the remaining dataset, the associated computational costs have driven the development of efficient, approximate unlearning techniques. Moving beyond data-centric MU approaches, our study introduces a novel model-based perspective: model sparsification via weight pruning, which is capable of reducing the gap between exact unlearning and approximate unlearning. We show in both theory and practice that model sparsity can boost the multi-criteria unlearning performance of an approximate unlearner, closing the approximation gap, while continuing to be efficient. This leads to a new MU paradigm, termed prune first, then unlearn, which infuses a sparse model prior into the unlearning process. Building on this insight, we also develop a sparsity-aware unlearning method that utilizes sparsity regularization to enhance the training process of approximate unlearning. Extensive experiments show that our proposals consistently benefit MU in various unlearning scenarios. A notable highlight is the 77% unlearning efficacy gain of fine-tuning (one of the simplest unlearning methods) when using sparsity-aware unlearning. Furthermore, we demonstrate the practical impact of our proposed MU methods in addressing other machine learning challenges, such as defending against backdoor attacks and enhancing transfer learning. Codes are available at https://github.com/OPTML-Group/Unlearn-Sparse.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (78)
  1. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy, pages 463–480. IEEE, 2015.
  2. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141–159. IEEE, 2021.
  3. A survey of machine unlearning. arXiv preprint arXiv:2209.02299, 2022.
  4. Jeffrey Rosen. The right to be forgotten. Stan. L. Rev. Online, 64:88, 2011.
  5. The european union general data protection regulation: what it is and what it means. Information & Communications Technology Law, 28(1):65–98, 2019.
  6. Machine unlearning of features and labels. arXiv preprint arXiv:2108.11577, 2021.
  7. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11516–11524, 2021.
  8. Unrolling sgd: Understanding factors influencing machine unlearning. arXiv preprint arXiv:2109.13398, 2021.
  9. Evaluating machine unlearning via epistemic uncertainty. arXiv preprint arXiv:2208.10836, 2022.
  10. Approximate data deletion from machine learning models. In International Conference on Artificial Intelligence and Statistics, pages 2008–2016. PMLR, 2021.
  11. On the necessity of auditable algorithmic definitions for machine unlearning. In 31st USENIX Security Symposium (USENIX Security 22), pages 4007–4022, 2022.
  12. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304–9312, 2020.
  13. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.
  14. The lottery tickets hypothesis for supervised and self-supervised pre-training in computer vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16306–16316, 2021.
  15. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018.
  16. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259–3269. PMLR, 2020a.
  17. Sanity checks for lottery tickets: Does your winning ticket really win the jackpot? Advances in Neural Information Processing Systems, 34:12749–12760, 2021.
  18. Advancing model pruning via bi-level optimization. In Advances in Neural Information Processing Systems, 2022.
  19. What is the state of neural network pruning? Proceedings of machine learning and systems, 2:129–146, 2020.
  20. Hydra: Pruning adversarially robust neural networks. Advances in Neural Information Processing Systems, 33:19655–19666, 2020.
  21. Quarantine: Sparsity can uncover the trojan attack trigger for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 598–609, 2022a.
  22. A winning hand: Compressing deep networks can improve out-of-distribution robustness. Advances in Neural Information Processing Systems, 34:664–676, 2021.
  23. The effect of model compression on fairness in facial expression recognition. arXiv preprint arXiv:2201.01709, 2022.
  24. Can model compression improve nlp fairness. arXiv preprint arXiv:2201.08542, 2022.
  25. Leveraging sparse linear layers for debuggable deep networks. In International Conference on Machine Learning, pages 11205–11216. PMLR, 2021.
  26. Can you win everything with a lottery ticket? Transactions of Machine Learning Research, 2022b.
  27. Privacy-preserving learning via deep net pruning. arXiv preprint arXiv:2003.01876, 2020.
  28. Against membership inference attack: Pruning is all you need. arXiv preprint arXiv:2008.13578, 2020a.
  29. Federated unlearning via class-discriminative pruning. In Proceedings of the ACM Web Conference 2022, pages 622–632, 2022a.
  30. Learning with recoverable forgetting. In European Conference on Computer Vision, pages 87–103. Springer, 2022.
  31. What do compressed deep neural networks forget? arXiv preprint arXiv:1911.05248, 2019.
  32. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885–1894. PMLR, 2017.
  33. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54–71, 2019.
  34. Residuals and influence in regression. New York: Chapman and Hall, 1982.
  35. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019.
  36. Machine unlearning: A survey. ACM Computing Surveys, 56(1):1–36, 2023.
  37. Woodfisher: Efficient second-order approximation for neural network compression. Advances in Neural Information Processing Systems, 33:18098–18109, 2020.
  38. Privacy risks of securing machine learning models against adversarial examples. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 241–257, 2019.
  39. Privacy risk in machine learning: Analyzing the connection to overfitting. In 2018 IEEE 31st computer security foundations symposium (CSF), pages 268–282. IEEE, 2018.
  40. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in Neural Information Processing Systems, 33:6377–6389, 2020.
  41. Pruning neural networks at initialization: Why are we missing the mark? arXiv preprint arXiv:2009.08576, 2020b.
  42. Optimization with sparsity-inducing penalties. Foundations and Trends® in Machine Learning, 4(1):1–106, 2012.
  43. Learning multiple layers of features from tiny images. 2009.
  44. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
  45. Reading digits in natural images with unsupervised feature learning. 2011.
  46. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009.
  47. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
  48. Systematic evaluation of privacy risks of machine learning models. arXiv preprint arXiv:2003.10595, 2020.
  49. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017.
  50. Backdoor defense with machine unlearning. arXiv preprint arXiv:2201.09538, 2022a.
  51. A data-based perspective on transfer learning. arXiv preprint arXiv:2207.05739, 2022.
  52. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485–3492. IEEE, 2010.
  53. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012.
  54. Ffcv: Accelerating training by removing data bottlenecks. 2022.
  55. Making ai forget you: Data deletion in machine learning. Advances in neural information processing systems, 32, 2019.
  56. Descent-to-delete: Gradient-based methods for machine unlearning. In Algorithmic Learning Theory, pages 931–962. PMLR, 2021.
  57. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pages 4126–4142. PMLR, 2021.
  58. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075–18086, 2021.
  59. Our data, ourselves: Privacy via distributed noise generation. In Annual international conference on the theory and applications of cryptographic techniques, pages 486–503. Springer, 2006.
  60. The right to be forgotten in federated learning: An efficient realization with rapid retraining. arXiv preprint arXiv:2203.07320, 2022b.
  61. Graph unlearning. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 499–513, 2022c.
  62. Certified graph unlearning. arXiv preprint arXiv:2206.09140, 2022.
  63. Gnndelete: A general strategy for unlearning in graph neural networks. arXiv preprint arXiv:2302.13406, 2023.
  64. Hard to forget: Poisoning attacks on certified machine unlearning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7691–7700, 2022.
  65. Hidden poison: Machine unlearning enables camouflaged poisoning attacks. In NeurIPS ML Safety Workshop, 2022.
  66. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023.
  67. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023a.
  68. Fair infinitesimal jackknife: Mitigating the influence of biased training data points without refitting. In Advances in Neural Information Processing Systems, 2022.
  69. Understanding instance-level impact of fairness constraints. In International Conference on Machine Learning, pages 23114–23130. PMLR, 2022b.
  70. Coresets via bilevel optimization for continual learning and streaming. Advances in Neural Information Processing Systems, 33:14879–14890, 2020.
  71. Dataset pruning: Reducing training data by examining generalization influence. arXiv preprint arXiv:2205.09329, 2022.
  72. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018.
  73. Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376, 2020b.
  74. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018.
  75. Data-model-circuit tri-design for ultra-light video intelligence on edge devices. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 745–750, 2023b.
  76. Scalable differential privacy with sparse network finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5059–5068, 2021.
  77. A privacy-preserving-oriented dnn pruning and mobile acceleration framework. In Proceedings of the 2020 on Great Lakes Symposium on VLSI, pages 119–124, 2020.
  78. On differentiating parameterized argmin and argmax problems with application to bi-level optimization. arXiv preprint arXiv:1607.05447, 2016.
User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (8)
  1. Jinghan Jia (30 papers)
  2. Jiancheng Liu (19 papers)
  3. Parikshit Ram (43 papers)
  4. Yuguang Yao (24 papers)
  5. Gaowen Liu (60 papers)
  6. Yang Liu (2253 papers)
  7. Pranay Sharma (26 papers)
  8. Sijia Liu (204 papers)
Citations (68)
View on Semantic Scholar

Summary

Model Sparsity Can Simplify Machine Unlearning: A Study

The paper, "Model Sparsity Can Simplify Machine Unlearning," explores the prominent issue of machine unlearning within the broader context of data regulation requirements. Machine unlearning is essential for eliminating the impact of specific data examples from trained models, thereby aligning with data privacy regulations such as the GDPR. Although exact unlearning is achievable by retraining models on the remaining dataset post removal of specific data points, this process is computationally intensive. Therefore, the paper presents a novel approach leveraging model sparsification, specifically weight pruning, to enhance approximate unlearning, which is typically faster but less accurate.

In the paper, the authors propose a new paradigm termed "prune first, then unlearn." This approach introduces model sparsification as a precursor to the unlearning process, with the aim of reducing the performance gap between approximate and exact unlearning methods. Through comprehensive theoretical analysis and empirical evaluations, the authors indicate that model sparsity plays a pivotal role in augmenting unlearning efficacy across a variety of metrics, including unlearning accuracy, membership inference success rate, and remaining accuracy.

Key Numerical Results and Claims

The authors report significant numerical results endorsing their approach. Notably, they highlight a 77% increase in the unlearning efficacy of the fine-tuning method, which is considered one of the simplest approximate unlearning procedures. This improvement underscores the potential impact of integrating model sparsity within the unlearning framework.

Furthermore, the model sparsification approach demonstrates benefits beyond unlearning, including enhancing models' resistance to backdoor attacks and improving transfer-learning capabilities. Through extensive experiments across diverse datasets and model architectures, the paper shows how model sparsity not only reduces the computational demands associated with retraining but also closely approximates the results achieved through exact unlearning methodologies.

Practical and Theoretical Implications

From a practical perspective, the integration of model sparsification within the unlearning process holds significant promise for reducing computational loads while maintaining high unlearning accuracy. The implications of this research are particularly pertinent for environments requiring frequent compliance with data privacy laws, offering a more efficient pathway to achieve such compliance without compromising model performance.

Theoretically, the paper offers insights into the relationship between model sparsity and data influence on learning models. By formalizing the impact of sparsity via weight pruning on the unlearning process, the paper contributes to a deeper understanding of how sparse models can closely mirror dense model behaviors in various performance metrics. This insight could potentially steer future research towards exploring model sparsity as a foundational trait in developing more efficient unlearning mechanisms.

Speculations on Future AI Developments

The framework presented has the potential to shape future developments in artificial intelligence, particularly in sectors needing rapid adaptation to changing data privacy norms. As model complexity and the volume of training data continue to expand, approaches like model sparsity that aim to reduce computational burden while ensuring compliance and accuracy will likely gain traction.

Moreover, exploring the integration of sparsity within more complex model architectures beyond the scope of this paper could unlock further efficiencies, positioning model sparsification as a key feature in developing scalable, dependable, and regulation-compliant AI systems.

In conclusion, the paper offers a valuable contribution to the ongoing discourse on machine unlearning by presenting a structured, sparsity-driven approach that addresses key challenges associated with existing methods. As AI systems advance, the principles and methods expounded in this paper could serve as a foundation for more sophisticated, privacy-preserving learning frameworks.

File Document Download Save Streamline Icon: https://streamlinehq.com PDF File Document Download Save Streamline Icon: https://streamlinehq.com Markdown