Memorization in large-scale text-to-image diffusion models poses significant security and intellectual property risks, enabling adversarial attribute extraction and the unauthorized reproduction of sensitive or proprietary features. While conventional dememorization techniques, such as regularization and data filtering, limit overfitting to specific training examples, they fail to systematically prevent the internalization of prohibited concept-level features. Simply discarding all images containing a sensitive feature wastes invaluable training data, necessitating a method for selective learning at the concept level. We introduce a gradient projection method designed to enforce a stringent requirement of concept-level feature exclusion. Our defense operates during backpropagation by systematically identifying and excising training signals aligned with embeddings of prohibited attributes. Specifically, we project each gradient update onto the orthogonal complement of the sensitive feature's embedding space, thereby zeroing out its influence on the model's weights. Our method integrates seamlessly into standard diffusion model training pipelines and complements existing defenses. We analyze our method against an adversary aiming for feature extraction. In extensive experiments, we demonstrate that our framework drastically reduces memorization while rigorously preserving generation quality and semantic fidelity. By reframing memorization control as selective learning, our approach establishes a new paradigm for IP-safe and privacy-preserving generative AI.

Text-to-image (T2I) generative models such as Stable Diffusion and FLUX can synthesize realistic, high-quality images directly from textual prompts. The resulting image quality depends critically on well-crafted prompts that specify both subjects and stylistic modifiers, which have become valuable digital assets. However, the rising value and ubiquity of high-quality prompts expose them to security and intellectual-property risks. One key threat is the prompt stealing attack, i.e., the task of recovering the textual prompt that generated a given image. Prompt stealing enables unauthorized extraction and reuse of carefully engineered prompts, yet it can also support beneficial applications such as data attribution, model provenance analysis, and watermarking validation. Existing approaches often assume white-box gradient access, require large-scale labeled datasets for supervised training, or rely solely on captioning without explicit optimization, limiting their practicality and adaptability. To address these challenges, we propose PROMPTMINER, a black-box prompt stealing framework that decouples the task into two phases: (1) a reinforcement learning-based optimization phase to reconstruct the primary subject, and (2) a fuzzing-driven search phase to recover stylistic modifiers. Experiments across multiple datasets and diffusion backbones demonstrate that PROMPTMINER achieves superior results, with CLIP similarity up to 0.958 and textual alignment with SBERT up to 0.751, surpassing all baselines. Even when applied to in-the-wild images with unknown generators, it outperforms the strongest baseline by 7.5 percent in CLIP similarity, demonstrating better generalization. Finally, PROMPTMINER maintains strong performance under defensive perturbations, highlighting remarkable robustness. Code: https://github.com/aaFrostnova/PromptMiner