SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders

Researchers introduce SAEmnesia, a supervised sparse autoencoder framework that enables efficient concept unlearning in diffusion models by binding concepts to individual neurons. The method reduces computational overhead by 96.67% compared to existing approaches and achieves 9.22% improvement on benchmark tests, with demonstrated robustness against adversarial attacks.

SAEmnesia addresses a fundamental challenge in machine learning safety: the ability to selectively remove unwanted concepts from trained diffusion models without degrading overall performance. Traditional concept unlearning struggles because knowledge distributes across numerous latent features, requiring extensive computational resources to identify and remove. By enforcing one-to-one concept-neuron mappings through supervised training, SAEmnesia centralizes feature representation, creating an interpretable architecture where each concept occupies a single, identifiable neuron.

This advancement builds on growing concerns about controlling generative AI outputs, particularly regarding inappropriate content generation. The research demonstrates practical applications beyond academic interest—successfully suppressing nudity on benchmark datasets and maintaining model robustness when adversarial actors attempt to circumvent safety mechanisms. The method's scalability advantage proves especially significant for sequential unlearning scenarios, where removing multiple concepts typically compounds computational difficulty.

For AI developers and safety researchers, SAEmnesia offers substantial operational benefits. The 96.67% reduction in hyperparameter search dramatically lowers the barrier to implementing targeted concept removal, enabling smaller teams and organizations to implement safety controls without specialized computational infrastructure. This democratization of unlearning technology could accelerate responsible AI deployment across commercial applications.

The framework's success in adversarial robustness testing suggests maturation toward production-ready safety mechanisms. Future development will likely focus on extending SAEmnesia to larger models and exploring whether the approach generalizes across different model architectures. The availability of open-source implementation invites community validation and iteration, potentially establishing new standards for interpretable and controllable AI systems.

• →SAEmnesia reduces hyperparameter search burden by 96.67% compared to existing sparse autoencoder unlearning methods.
• →One-to-one concept-neuron mapping centralizes feature representation, enabling interpretable and targeted concept erasure.
• →Framework demonstrates 28.4% accuracy improvement in sequential unlearning scenarios with nine objects removed.
• →Method proves robust against adversarial attacks while effectively suppressing unwanted content like nudity.
• →Open-source availability enables broader adoption of interpretable concept unlearning across AI development communities.