What makes unlearning hard and what to do about it

TL;DR

This paper identifies two key factors affecting unlearning difficulty and the performance of unlearning algorithms and develops a framework coined Refined-Unlearning Meta-algorithm (RUM), a meta-algorithm that employs existing algorithms to unlearn each subset and finally delivers a model that has unlearned the overall forget set.

Abstract

Machine unlearning is the problem of removing the effect of a subset of training data (the ''forget set'') from a trained model without damaging the model's utility e.g. to comply with users' requests to delete their data, or remove mislabeled, poisoned or otherwise problematic data. With unlearning research still being at its infancy, many fundamental open questions exist: Are there interpretable characteristics of forget sets that substantially affect the difficulty of the problem? How do these characteristics affect different state-of-the-art algorithms? With this paper, we present the first investigation aiming to answer these questions. We identify two key factors affecting unlearning difficulty and the performance of unlearning algorithms. Evaluation on forget sets that isolate these identified factors reveals previously-unknown behaviours of state-of-the-art algorithms that don't materialize on random forget sets. Based on our insights, we develop a framework coined Refined-Unlearning Meta-algorithm (RUM) that encompasses: (i) refining the forget set into homogenized subsets, according to different characteristics; and (ii) a meta-algorithm that employs existing algorithms to unlearn each subset and finally delivers a model that has unlearned the overall forget set. We find that RUM substantially improves top-performing unlearning algorithms. Overall, we view our work as an important step in (i) deepening our scientific understanding of unlearning and (ii) revealing new pathways to improving the state-of-the-art.

Figures (15)

• Figure 1: Uncovering two factors that affect unlearning difficulty according to ToW (where higher is better). Left: the more entangled the retain and forget sets are in the embedding space, the harder it is to unlearn. Right: the less memorized a forget set is (thus having influenced the model less), the easier it is to unlearn (for most algorithms). Error bars correspond to 95% confidence intervals from running each algorithm 3 times (6 times for relabelling-based that had higher variance).
• Figure 2: Overview of RUM.
• Figure 3: From subplots a and c, we observe that RUM$^\mathcal{F}$ improves each unlearning algorithm. Vanilla corresponds to unlearning $\mathcal{S}$ in one go, whereas Shuffle and RUM$^\mathcal{F}$ operate sequentially on 3 subsets of $\mathcal{S}$. In the case of RUM$^\mathcal{F}$, the 3 subsets are the result of applying $\mathcal{F}$ and the order is low $\rightarrow$ medium $\rightarrow$ high, whereas Shuffle uses equal-sized random subsets, serving as a control experiment. Further, from subplot b, we observe that full RUM, equipped with the best algorithm for each subset (do nothing $\rightarrow$ Fine-tune $\rightarrow$ SalUn), yields the overall best results (note: in CIFAR-100, NegGrad+ is the best algorithm, so full RUM corresponds to the RUM$^\mathcal{F}$ variant of NegGrad+).
• Figure 4: Replacing mem scores with the efficient C-proxy yields similar trends and performance gains, carving a path for practical deployment of RUM. Left: forget sets with lower C-proxy values (i.e., higher mem scores, since C-proxy and memorization are negatively correlated) are harder to unlearn, consistent with the trend in Figure \ref{['fig:mem-vs-tow']}. Right: RUM$^\mathcal{F}$ using C-proxy in the refinement step enhances unlearning performance across algorithms, comparable to using the memorization score in Figure \ref{['fig:rumf-tow-cifar10']}. Error bars correspond to 95% confidence intervals, with each algorithm run 3 times.
• Figure 5: Sequence dynamics for SalUn RUM$^\mathcal{F}$ on CIFAR-10. We report the accuracy on overall $\mathcal{S}$, $\mathcal{R}$ and $\mathcal{D}_{test}$, and subsets of $\mathcal{S}$ after each step. Both orderings yield similar Tow (see Table \ref{['tab:rum-cifar10']}).

Theorems & Definitions (2)

• Definition 2.1
• Definition 2.2