This paper introduces a more rigorous way to audit "concept erasure" in text-to-image diffusion models. Instead of relying on "prompt attacks" that only test the model's interface, the authors propose least-action auditing, which measures the model's internal generative capacity to produce a forbidden concept. This approach not only predicts real-world "leakage" better than existing methods but also provides a mechanism for repairing models that have been incompletely erased.
TL;DR
The authors move beyond simple prompt-based testing to audit concept erasure by measuring the "least-action barrier"—the minimum mathematical effort required to force a model to generate an erased concept. This metric accurately predicts real-world concept leakage and enables a new, automated way to "patch" models to ensure concepts are more deeply removed.
What Problem Does This Paper Address?
When we try to "erase" a concept (like a specific artistic style or NSFW content) from a diffusion model, we usually audit the result with prompt attacks. We try to find a specific text string that still triggers the concept.
The problem is that a model might refuse a direct prompt but still have the concept "hidden" in its internal generative dynamics. A model that refuses the prompt "a photo of a nude person" might still be capable of generating that content if steered slightly differently. Current auditing methods only test the interface (the prompt), not the dynamics (the score field). This paper addresses the need for a "model-internal" diagnosis of residual generative capacity.
Methods
The core idea is to treat auditing as a stochastic control problem.
- Least-Action Barrier ($\Delta S_{\text{rob}}$): The authors add a "control" signal to the noise channel of a frozen diffusion model. They then use optimization (Adam) to find the "path of least resistance"—the minimum control energy needed to force the model to generate the target concept. This energy is the "barrier." A high barrier means the concept is deeply removed; a low barrier means it’s still easily accessible.
- Detector-Consensus Terminal: To ensure the audit is honest and hasn't just "gamed" a single detector, they use a consensus of two different AI models (CLIP and DINOv2). Both must agree that the generated image contains the target concept for the audit to be considered successful.
- Audit-Guided Repair: The "least-action" paths found during the audit (called "instantons") represent the model's remaining "leak modes." The authors use these modes to create a closed-form mathematical patch that projects these residual capacities out of the model's weights.
Datasets
The experiments were conducted on Stable Diffusion v1.4.
- Concepts: The audit covered 18 concept-mechanism pairs, including 8 artistic styles, 5 ImageNette objects (e.g., church, parachute, gas pump), and NSFW nudity.
- Benchmarks: Real-world "leakage" was measured using NudeNet (for NSFW content) and ImageNet classifiers (for objects) on 1,500 fresh samples per model. The I2P benchmark was used to test sensitivity to adversarial prompts.
Key Insights and Ideas
- Interface vs. Dynamics: Some erasure methods (like Safe Latent Diffusion) only suppress the prompt interface but leave the concept perfectly intact in the model's internal "score space." These models have a near-zero barrier and high leakage under pressure.
- The "Where" of Concepts: The audit reveals that concepts typically live in the early, high-noise steps of the generation process. Removing the concept from these early steps is crucial for effective erasure.
- Barrier as Probability: Mathematically, the least-action barrier is equivalent to the negative log-probability of the concept occurring. This grounds the audit in solid probability theory rather than just empirical "red-teaming."
- Repairability: The audit doesn't just find holes; it provides the exact directions needed to plug them. Audit-guided repairs were found to be significantly more effective than "blind" edits.
Key Results
- Predictive Power: The barrier $\Delta S_{\text{rob}}$ is a strong predictor of actual deployment leakage (pooled Spearman $\rho = -0.76$). In every single case (18/18), the model with the higher barrier leaked less.
- Eraser Rankings: The audit successfully ranked common erasure methods by their "depth." For NSFW nudity: UCE (low barrier, 16.4% leak) < ESD (75.5 barrier, 9.8% leak) < RECE (224.5 barrier, 1.4% leak).
- Effective Repair: Using the audit to guide a patch reduced UCE nudity leakage from 16.2% to 0/1,500 samples, without damaging the model's ability to generate unrelated content.
Limitations and Future Work
- Detector Dependency: The method requires a differentiable "indicator" (like CLIP) and a consensus detector. If no good detector exists for a concept, the audit cannot be performed.
- Backbone Specificity: While the authors showed the audit transfers to a distilled backbone (BK-SDM), it is currently designed for latent diffusion models. Extending it to newer architectures like rectified flow is a suggested theoretical next step.
- Computational Cost: A full audit takes 1–5 GPU-minutes per model, which is efficient but still more intensive than simple prompt testing.
Summary updated