Representation Unlearning: Forgetting through Information Compression

TL;DR

Representation Unlearning is introduced, a framework that performs unlearning directly in the model's representation space and achieves more reliable forgetting, better utility retention, and greater computational efficiency than parameter-centric baselines.

Abstract

Machine unlearning seeks to remove the influence of specific training data from a model, a need driven by privacy regulations and robustness concerns. Existing approaches typically modify model parameters, but such updates can be unstable, computationally costly, and limited by local approximations. We introduce Representation Unlearning, a framework that performs unlearning directly in the model's representation space. Instead of modifying model parameters, we learn a transformation over representations that imposes an information bottleneck: maximizing mutual information with retained data while suppressing information about data to be forgotten. We derive variational surrogates that make this objective tractable and show how they can be instantiated in two practical regimes: when both retain and forget data are available, and in a zero-shot setting where only forget data can be accessed. Experiments across several benchmarks demonstrate that Representation Unlearning achieves more reliable forgetting, better utility retention, and greater computational efficiency than parameter-centric baselines.

Figures (45)

• Figure 1: The model is first trained on both the retain set $\mathcal{D}_r$ (blue) and the forget set $\mathcal{D}_f$ (red), leading the learned representation $Z$ to encode information from both. Representation Unlearning seeks to learn a transformation $f_\phi$ that maps $Z$ to a new representation $Z'$ in which information attributable to $\mathcal{D}_f$ is removed while information relevant to $\mathcal{D}_r$ is preserved.
• Figure 2: $I(Z'; Z \mid X_r) > 0$ indicates that the transformation $f_\phi$ removes some information relevant to $X_r$. Conversely, $I(Z'; X_f) > 0$ indicates that $f_\phi$ does not fully eliminate information associated with $X_f$.
• Figure 3: Original Model
• Figure 4: Standard Unlearning (w/ $\mathcal{D}_r$)
• Figure 5: Zero-Shot Unlearning (w/o $\mathcal{D}_r$)