SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models
Researchers Nicola Novello and Andrea M. Tonello have introduced SPACE (SParse cross-Attention-based Concept Erasure), a novel method designed to erase specific concepts from text-to-image diffusion models. This technique addresses the growing need to prevent the generation of copyrighted or explicit content, a challenge that becomes increasingly difficult as models scale from smaller architectures like Stable Diffusion 1.5 to larger ones such as Stable Diffusion XL. Unlike traditional backpropagation-based techniques, SPACE utilizes a closed-form update to iteratively modify cross-attention parameters, jointly inducing sparsity and erasing target concepts. By concentrating concept mapping into a lower-dimensional subspace, the method achieves superior erasure efficacy and robustness against adversarial prompts compared to dense baselines. Experimental results demonstrate that SPACE attains 80%-90% cross-attention sparsity, which significantly reduces storage requirements for modified parameters by 70%. This advancement offers a memory-efficient and effective solution for maintaining safety and compliance in large-scale generative AI systems, marking a significant step forward in the field of machine learning and artificial intelligence security.
Wire timeline
SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models
Researchers Nicola Novello and Andrea M. Tonello have introduced SPACE (SParse cross-Attention-based Concept Erasure), a novel method designed to erase specific concepts from text-to-image diffusion models. This technique addresses the growing need to prevent the generation of copyrighted or explicit content, a challenge that becomes increasingly difficult as models scale from smaller architectures like Stable Diffusion 1.5 to larger ones such as Stable Diffusion XL. Unlike traditional backpropagation-based techniques, SPACE utilizes a closed-form update to iteratively modify cross-attention parameters, jointly inducing sparsity and erasing target concepts. By concentrating concept mapping into a lower-dimensional subspace, the method achieves superior erasure efficacy and robustness against adversarial prompts compared to dense baselines. Experimental results demonstrate that SPACE attains 80%-90% cross-attention sparsity, which significantly reduces storage requirements for modified parameters by 70%. This advancement offers a memory-efficient and effective solution for maintaining safety and compliance in large-scale generative AI systems, marking a significant step forward in the field of machine learning and artificial intelligence security.
cs.AI updates on arXiv.org