Exact Machine Unlearning on Weightless Neural Networks

• Luan Martins UFRJ
• Felipe M. G. França UFRJ
• Priscila Machado Vieira Lima UFRJ
• Claudio Miceli de Farias UFRJ

DOI: https://doi.org/10.5753/eniac.2026.30819

Resumo

Privacy regulations create pressure to remove personal records and, when those records have influenced a trained model, motivate machine unlearning. Existing methods make a difficult trade-off: retraining is exact but costly, whereas inexpensive approximations may leave detectable information about deleted examples. We examine weightless neural networks of the WiSARD family, whose training procedure adds independent integer-counter updates. Reversing those updates returns the counters to the state produced without the deleted examples, which permits a cell-by-cell exactness certificate. We evaluate this mechanism on a calibrated synthetic Where’s-Waldo benchmark against full and sharded retraining, deletion-enabled forests, and gradient ascent. In class erasure, WiSARD deleted the selected class in 40 ms, retained 0.899 accuracy, used 9.3 MB, produced a maximum counter difference of zero, and reduced membership-inference area under the ROC curve to 0.50; retraining required 1.42 s, while the more accurate deletion-enabled forest retained 0.953 accuracy at 13.87 s and 1.59 GB. Within the evaluated setting, WiSARD therefore provides an exact option for applications that prioritize deletion latency, although it does not provide the highest accuracy or the smallest absolute memory footprint.

Referências

Aleksander, I., Thomas, W. V., and Bowden, P. A. (1984). WISARD: A radical step forward in image recognition. Sensor Review, 4(3):120–124.

Barthelmes, T. and Vidal, F. P. (2021). Where’s wally? a machine learning approach. In Computer Graphics and Visual Computing (CGVC), pages 28–31. The Eurographics Association.

Bledsoe, W. W. and Browning, I. (1959). Pattern recognition and reading by machine. In Papers Presented at the Eastern Joint IRE-AIEE-ACM Computer Conference, pages 225–232. ACM.

Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C. A., Jia, H., Travers, A., Zhang, B., Lie, D., and Papernot, N. (2021). Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141–159. IEEE.

Brasil (2018). Lei n. 13.709/2018: Lei geral de proteção de dados pessoais (LGPD). Diário Oficial da União, Brasília, DF. Artigo 18: direito de eliminação.

Brophy, J. and Lowd, D. (2021). Machine unlearning for random forests. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Research, pages 1092–1104. PMLR.

Cao, Y. and Yang, J. (2015). Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463–480. IEEE.

Carlini, N., Chien, S., Nasr, M., Song, S., Terzis, A., and Tramèr, F. (2022). Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897–1914. IEEE.

Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., and Zhang, Y. (2021). When machine unlearning jeopardizes privacy. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 896–911. ACM.

Constantinou, V. (2017). Hey, Waldo: Labeled images of the Where’s Waldo puzzle. [link]. GitHub repository.

European Parliament and Council of the European Union (2016). Regulation (EU) 2016/679 (general data protection regulation). Official Journal of the European Union, L 119. Article 17: Right to erasure (‘right to be forgotten’).

Ginart, A., Guan, M., Valiant, G., and Zou, J. (2019). Making AI forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.

Hayes, J., Shumailov, I., Triantafillou, E., Khalifa, A., and Papernot, N. (2025). Inexact unlearning needs more careful evaluations to avoid a false sense of privacy. In 2025 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 497–519. IEEE.

Laguna, S., da Silva Gonçalves, J., Vandenhirtz, M., Ryser, A., Cannistraci, I., and Vogt, J. E. (2026). Rethinking machine unlearning: Models designed to forget via key deletion. In ICLR 2026 Workshop on Test-Time Updates (TTU). Oral presentation.

Lima Filho, A. S., Guarisa, G. P., Lusquino Filho, L. A. D., Oliveira, L. F. R., França, F. M. G., and Lima, P. M. V. (2020). wisardpkg: A library for WiSARD-based models. arXiv preprint arXiv:2005.00887.

Lusquino Filho, L. A. D., Oliveira, L. F. R., Lima Filho, A. S., Guarisa, G. P., Felix, L. M., Lima, P. M. V., and França, F. M. G. (2020). Extending the weightless WiSARD classifier for regression. Neurocomputing, 416:280–291.

Mavrothalassitis, I., Puigdemont, P., Levi, N., and Cevher, V. (2025). Ascent fails to forget. In Advances in Neural Information Processing Systems, volume 38, pages 45330–45365. Curran Associates, Inc.

Nguyen, T. T., Huynh, T. T., Ren, Z., Nguyen, P. L., Liew, A. W.-C., Yin, H., and Nguyen, Q. V. H. (2025). A survey of machine unlearning. ACM Transactions on Intelligent Systems and Technology, 16(5):108:1–108:46.

Shokri, R., Stronati, M., Song, C., and Shmatikov, V. (2017). Membership inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy (SP), pages 3–18. IEEE.

Thudi, A., Jia, H., Shumailov, I., and Papernot, N. (2022). On the necessity of auditable algorithmic definitions for machine unlearning. In 31st USENIX Security Symposium (USENIX Security 22), pages 4007–4022. USENIX Association.

Xu, H., Zhu, T., Zhang, L., Zhou, W., and Yu, P. S. (2023). Machine unlearning: A survey. ACM Computing Surveys, 56(1):1–36.