Abstract
We show that the influence of a subset of the training samples can be removed -- or "forgotten" -- from the weights of a network trained on large-scale image classification tasks, and we provide strong computable bounds on the amount of remaining information after forgetting. Inspired by real-world applications of forgetting techniques, we introduce a novel notion of forgetting in mixed-privacy setting, where we know that a "core" subset of the training samples does not need to be forgotten. While this variation of the problem is conceptually simple, we show that working in this setting significantly improves the accuracy and guarantees of forgetting methods applied to vision classification tasks. Moreover, our method allows efficient removal of all information contained in non-core data by simply setting to zero a subset of the weights with minimal loss in performance. We achieve these results by replacing a standard deep network with a suitable linear approximation. With opportune changes to the network architecture and training procedure, we show that such linear approximation achieves comparable performance to the original network and that the forgetting problem becomes quadratic and can be solved efficiently even for large models. Unlike previous forgetting methods on deep networks, ours can achieve close to the state-of-the-art accuracy on large scale vision tasks. In particular, we show that our method allows forgetting without having to trade off the model accuracy.
Introduction
In the era of data privacy regulations such as GDPR, the "right to be forgotten" presents a significant technical challenge for machine learning. Simply deleting a user's data from a database does not remove its influence, which remains encoded within the model's weights. While retraining a model from scratch is the most thorough way to ensure data is forgotten, it is prohibitively expensive for modern deep neural networks (DNNs).
Previous attempts at "machine unlearning" have often been limited to simple linear models or have struggled with the non-convex loss landscapes of deep networks, where small data changes can lead to unpredictable weight shifts. We address these challenges by introducing Mixed-Linear Forgetting (ML-Forgetting), a framework that enables efficient and verifiable data removal in large-scale vision tasks.
The Mixed-Privacy Framework
Our approach hinges on a "mixed-privacy" setting. We categorize training data into two groups:
- Core Data ($D_c$): Generic, stable data (e.g., ImageNet) that is not subject to deletion requests.
- User Data ($D$): Task-specific or sensitive data that may need to be forgotten.
By linearizing the network around the core weights using a first-order Taylor expansion, we transform the complex optimization landscape into a quadratic one. In this linearized regime, forgetting becomes a mathematically tractable problem. We can remove the influence of specific samples using a "reverse Newton step" and ensure complete information removal through a calculated "scrubbing" noise process.
Experimental Results
Accuracy and Performance
A key concern with linearization is the potential loss of predictive power. However, our results demonstrate that with specific architectural adjustments—such as using Leaky ReLU activations—the ML-Forgetting model achieves test accuracy comparable to standard non-linear DNNs across various benchmarks, including Caltech-256 and fine-grained datasets like FGVC Aircrafts.
As shown in the figure above, the Mixed-Linear approach consistently matches or nears the performance of standard DNNs, significantly outperforming simpler baselines like training only the last layer.
Efficiency and Information Removal
The primary advantage of ML-Forgetting is its speed. While retraining from scratch typically requires approximately 50 epochs, our forgetting process achieves its goals in just 2–3 passes over the remaining data. This represents a 16–25x speedup.
The figure above illustrates that the information content (measured in MNATS) drops precipitously within the first few epochs of the forgetting process, while the test error remains remarkably stable. This confirms that we can rapidly "scrub" sensitive information without degrading the model's utility.
Conclusion
ML-Forgetting bridges the gap between the theoretical requirements of data privacy and the practical constraints of deep learning. By leveraging a mixed-privacy setting and linear approximations, we provide a method that is not only computationally efficient but also offers strong, computable guarantees that data has been truly forgotten. This framework allows deep networks to maintain state-of-the-art accuracy while respecting the dynamic privacy needs of users.
Summary updated