This study investigates the generation of unsafe or harmful content in state-of-the-art generative models, focusing on methods for restricting such generations. We introduce a novel training-free approach using attention reweighing to remove unsafe concepts without additional training during inference. We compare our method against existing ablation methods, evaluating the performance on both, direct and adversarial jailbreak prompts, using qualitative and quantitative metrics. We hypothesize potential reasons for the observed results and discuss the limitations and broader implications of content restriction.
Conditioned Activation Transport for T2I Safety Steering
NASK National Research Institute(NASK国家研究所);Warsaw University of Technology(华沙技术大学);IDEAS Research Institute(IDEAS研究 institute);CISPA Helmholtz Center for Information Security(CISPA海德堡中心 for 信息安全)
Erased, But Not Forgotten: Erased Rectified Flow Transformers Still Remain Unsafe Under Concept Attack
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University(北京航空航天大学人工智能学院未来区块链与隐私计算北京高精尖创新中心);University of Science and Technology of China(中国科学技术大学);National University of Defense Technology(国防科技大学);GigaAI(极佳科技);National University of Singapore(新加坡国立大学)
NOUS: Video-Driven 3D Human Reaction Generation via Observation-Reaction Mutual Steering
Nanyang Technological University(南洋理工大学);National University of Singapore(新加坡国立大学);Renmin University of China(中国人民大学);University of Science and Technology of China(中国科学技术大学);Hefei University of Technology(合肥工业大学)
When Safety Collides: Resolving Multi-Category Harmful Conflicts in Text-to-Image Diffusion via Adaptive Safety Guidance
Sydney AI Centre, The University of Sydney(悉尼人工智能中心,悉尼大学);City University of Hong Kong(香港城市大学);TMLR Group, Hong Kong Baptist University(Baptist大学TMLR小组)