TLDR: A new research paper reveals that common machine unlearning methods, particularly those based on gradient ascent (Descent-Ascent), frequently fail because they overlook the statistical connections between the data intended for removal (forget set) and the data meant to be kept (retain set). This interdependence can cause unlearning algorithms to degrade model performance, diverge from ideal solutions, and even make the model worse than its original state, highlighting a fundamental flaw in these widely used techniques.
Machine learning models have become indispensable in various fields, from medicine to cybersecurity. However, with their widespread use comes the critical need for data privacy and model maintenance. This is where ‘machine unlearning’ comes into play – the process of selectively removing the influence of specific training examples from a model without compromising its overall performance.
Traditionally, if you wanted a model to ‘forget’ certain data, the most straightforward (though computationally expensive) approach was to retrain the model from scratch on only the data you wished to keep. As unlearning requests become more frequent and datasets grow larger, this ‘retrain from scratch’ method becomes impractical. This has led to the development of more efficient, approximate unlearning techniques.
Many popular unlearning methods, especially in deep learning, rely on what are called Descent-Ascent (DA) algorithms. These methods attempt to reverse the impact of the ‘forget set’ (the data to be removed) by performing gradient ascent on these points, while simultaneously performing gradient descent on the ‘retain set’ (the data to be kept). The idea is to push the model away from the forgotten data’s influence.
However, a recent research paper titled ‘Ascent Fails to Forget’ by Ioannis Mavrothalassitis, Pol Puigdemont, Noam Itzhak Levi, and Volkan Cevher from EPFL, challenges the effectiveness of these DA-based methods. Contrary to common assumptions, the researchers demonstrate that these gradient ascent-based techniques frequently fail to perform machine unlearning effectively. The core reason for this failure, they argue, is an often-overlooked factor: the inherent statistical dependence between the forget and retain datasets.
The Overlooked Connection: Data Dependencies
The paper highlights that even simple correlations between the data you want to forget and the data you want to keep can undermine the entire unlearning process. The common belief is that these two sets can be manipulated independently during unlearning. However, the researchers provide both empirical and theoretical evidence showing that DA methods often fail precisely because of this unaddressed relationship.
For instance, if the ‘forget set’ is chosen randomly from the original dataset, it will naturally share high statistical dependence with the ‘retain set’. In such cases, trying to degrade the model’s performance on the forget set (a goal of many DA methods) inevitably harms the model’s overall performance on new, unseen data. This means that instead of truly ‘unlearning’, the model simply becomes worse.
Empirical and Theoretical Insights
The researchers conducted experiments using ResNet-9 models on the Cifar-10 dataset, evaluating different gradient-based unlearning approaches like Gradient Ascent (GA) and Gradient Descent/Ascent (GDA). They used a metric called KLoM (KL Divergence of Margins), where a score close to zero indicates perfect unlearning. Their findings were stark: both GA and GDA methods either failed to significantly move away from the original model’s state or severely degraded its performance.
A critical observation was termed the ‘Ascent Forgets Illusion’. It showed that by extensively tuning hyperparameters and selecting the ‘best’ run, one could be tricked into believing unlearning was successful, even with simple GA. This illusion stems from the difficulty of knowing when to stop the unlearning process, as different data points ‘unlearn’ at different rates.
The theoretical analysis further solidified these empirical observations. In simplified models like logistic regression, the paper demonstrates that inter-set correlations cause DA iterations to progressively diverge from the ‘oracle’ – the ideal model trained only on the retain set. Strikingly, these methods can converge to solutions that are not only far from the oracle but potentially even further from it than the original model itself, making the unlearning process actively detrimental.
A simple example illustrated how these dependencies can trap models in inferior local minima, which cannot be escaped even with further fine-tuning. This means that once a model is ‘damaged’ by DA unlearning, it might be irreversibly stuck in a suboptimal state.
Also Read:
- Verifying Data Removal in AI Models with Self-supervised Seeding
- Unraveling Agentic RL Failures: Beyond Reward Miscalibration
Implications and Future Directions
The findings of ‘Ascent Fails to Forget’ have significant implications for the field of machine unlearning. They highlight that the presence of statistical dependencies, even subtle correlations, is sufficient for ascent-based unlearning to fail. This suggests that current DA methods, which disregard these dependencies, are fundamentally flawed for reliable and consistent machine unlearning in real-world scenarios.
The authors suggest that future research on ascent-based methods must explicitly consider these data dependencies. They also point to alternative unlearning strategies, such as methods based on ‘rewinding’ or stochastic approaches involving noise, as potentially more robust alternatives when the properties of the dataset are unknown.
This research serves as a crucial warning for practitioners relying on gradient ascent-based unlearning. It underscores the need for more rigorous evaluation and the development of new algorithms that can genuinely achieve machine unlearning without inadvertently harming model performance or creating privacy paradoxes. You can read the full paper here: Ascent Fails to Forget.
