Reference-Specific Unlearning Metrics Can Hide the Truth: A Reality Check

TL;DR

This work addresses a core problem in machine unlearning: ensuring that an unlearned model is truly indistinguishable from a retain-only model trained without the forgotten data. It introduces FADE, a distributional alignment metric that compares bidirectional likelihoods between the unlearned and retain-only models, making it modality-agnostic for both language and vision. Across TOFU (LLMs) and UnlearnCanvas (diffusion models), FADE reveals that many methods achieving high scores on traditional, reference-based metrics fail to achieve true distributional equivalence, and can even drift further from the retain-only oracle. By providing a robust, distributional evaluation framework, FADE offers a principled basis for developing and validating genuinely effective unlearning methods with stronger resistance to post-unlearning attacks.

Abstract

Current unlearning metrics for generative models evaluate success based on reference responses or classifier outputs rather than assessing the core objective: whether the unlearned model behaves indistinguishably from a model that never saw the unwanted data. This reference-specific approach creates systematic blind spots, allowing models to appear successful while retaining unwanted knowledge accessible through alternative prompts or attacks. We address these limitations by proposing Functional Alignment for Distributional Equivalence (FADE), a novel metric that measures distributional similarity between unlearned and reference models by comparing bidirectional likelihood assignments over generated samples. Unlike existing approaches that rely on predetermined references, FADE captures functional alignment across the entire output distribution, providing a principled assessment of genuine unlearning. Our experiments on the TOFU benchmark for LLM unlearning and the UnlearnCanvas benchmark for text-to-image diffusion model unlearning reveal that methods achieving near-optimal scores on traditional metrics fail to achieve distributional equivalence, with many becoming more distant from the gold standard than before unlearning. These findings expose fundamental gaps in current evaluation practices and demonstrate that FADE provides a more robust foundation for developing and assessing truly effective unlearning methods.

Figures (8)

• Figure 1: Illustration of Functional Alignment for Distributional Equivalence (FADE), our proposed metric for unlearning efficacy. FADE measures the distributional distance between the retain-only model and the unlearned model based on conditional output distributions of the two models.
• Figure 2: NLL distributions from the unlearned model (y-axis) and the retain-only model (x-axis). Each dot represents a single sample from $\mathcal{D}_{\text{forget}}$. Each plot shows results from using paraphrased answers or original answers for NLL computation. The forget quality depends significantly on which reference answer is used, as the NLL distributions heavily depend on which answers are used.
• Figure 3: Example outputs given questions from $\mathcal{D}_{\text{forget}}$ after unlearning 1% of TOFU via Gradient Ascent (GA). While the improvement in forget quality value from the base model's -20.73 implies successful unlearning, the actual degree of forgetting is on the opposite extreme, with the unlearned model completely recovering the original answers.
• Figure 4: Example images generated by prompting various models to generate Van Gogh style images.
• Figure 5: FADE (top row; lower is better) and FQ (bottom; y-axis is inverted, lower is better) measurements from unlearning varying forget sets in the TOFU benchmark. The unlearning methods tested are GA, GD, NPO, DPO, and IHL. The dashed line in FADE represents the baseline FADE value due to randomness in training. The dashed line in FQ is the optimal value at zero. The shaded regions indicate the standard deviation across 3 random seeds.