Safety Alignment Backfires: Preventing the Re-emergence of Suppressed Concepts in Fine-tuned Text-to-Image Diffusion Models

• URL: http://arxiv.org/abs/2412.00357v1
• Date: Sat, 30 Nov 2024 04:37:38 GMT
• Title: Safety Alignment Backfires: Preventing the Re-emergence of Suppressed Concepts in Fine-tuned Text-to-Image Diffusion Models
• Authors: Sanghyun Kim, Moonseok Choi, Jinwoo Shin, Juho Lee,
• Abstract summary: Fine-tuning text-to-image diffusion models can inadvertently undo safety measures, causing models to relearn harmful concepts. We present a novel but immediate solution called Modular LoRA, which involves training Safety Low-Rank Adaptation modules separately from Fine-Tuning LoRA components. This method effectively prevents the re-learning of harmful content without compromising the model's performance on new tasks.
• Score: 57.16056181201623
• License:
• Abstract: Fine-tuning text-to-image diffusion models is widely used for personalization and adaptation for new domains. In this paper, we identify a critical vulnerability of fine-tuning: safety alignment methods designed to filter harmful content (e.g., nudity) can break down during fine-tuning, allowing previously suppressed content to resurface, even when using benign datasets. While this "fine-tuning jailbreaking" issue is known in large language models, it remains largely unexplored in text-to-image diffusion models. Our investigation reveals that standard fine-tuning can inadvertently undo safety measures, causing models to relearn harmful concepts that were previously removed and even exacerbate harmful behaviors. To address this issue, we present a novel but immediate solution called Modular LoRA, which involves training Safety Low-Rank Adaptation (LoRA) modules separately from Fine-Tuning LoRA components and merging them during inference. This method effectively prevents the re-learning of harmful content without compromising the model's performance on new tasks. Our experiments demonstrate that Modular LoRA outperforms traditional fine-tuning methods in maintaining safety alignment, offering a practical approach for enhancing the security of text-to-image diffusion models against potential attacks.

Related papers

• NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning [37.024666077902225]
  A handful of malicious data uploaded by users can subtly manipulate the finetuning process, resulting in an alignment-broken model. Existing methods to counteract fine-tuning attacks typically require substantial computational resources. We propose textbfNeuron-textbfLevel textbfSafety textbfRealignment.
  arXiv  Detail & Related papers  (2024-12-17T02:59:04Z)
• Safety Without Semantic Disruptions: Editing-free Safe Image Generation via Context-preserving Dual Latent Reconstruction [49.60774626839712]
  Training multimodal generative models can expose users to harmful, unsafe and controversial or culturally-inappropriate outputs. We propose a modular, dynamic solution that leverages safety-context embeddings and a dual reconstruction process to generate safer images. We achieve state-of-the-art results on safe image generation benchmarks, while offering controllable variation of model safety.
  arXiv  Detail & Related papers  (2024-11-21T09:47:13Z)
• Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models [76.39651111467832]
  We introduce Reliable and Efficient Concept Erasure (RECE), a novel approach that modifies the model in 3 seconds without necessitating additional fine-tuning. To mitigate inappropriate content potentially represented by derived embeddings, RECE aligns them with harmless concepts in cross-attention layers. The derivation and erasure of new representation embeddings are conducted iteratively to achieve a thorough erasure of inappropriate concepts.
  arXiv  Detail & Related papers  (2024-07-17T08:04:28Z)
• Rethinking and Defending Protective Perturbation in Personalized Diffusion Models [21.30373461975769]
  We study the fine-tuning process of personalized diffusion models (PDMs) through the lens of shortcut learning. PDMs are susceptible to minor adversarial perturbations, leading to significant degradation when fine-tuned on corrupted datasets. We propose a systematic defense framework that includes data purification and contrastive decoupling learning.
  arXiv  Detail & Related papers  (2024-06-27T07:14:14Z)
• Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models? [52.238883592674696]
  Ring-A-Bell is a model-agnostic red-teaming tool for T2I diffusion models. It identifies problematic prompts for diffusion models with the corresponding generation of inappropriate content. Our results show that Ring-A-Bell, by manipulating safe prompting benchmarks, can transform prompts that were originally regarded as safe to evade existing safety mechanisms.
  arXiv  Detail & Related papers  (2023-10-16T02:11:20Z)
• Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts [63.61248884015162]
  Text-to-image diffusion models have shown remarkable ability in high-quality content generation. This work proposes Prompting4 Debugging (P4D) as a tool that automatically finds problematic prompts for diffusion models. Our result shows that around half of prompts in existing safe prompting benchmarks which were originally considered "safe" can actually be manipulated to bypass many deployed safety mechanisms.
  arXiv  Detail & Related papers  (2023-09-12T11:19:36Z)
• Towards Safe Self-Distillation of Internet-Scale Text-to-Image Diffusion Models [63.20512617502273]
  We propose a method called SDD to prevent problematic content generation in text-to-image diffusion models. Our method eliminates a much greater proportion of harmful content from the generated images without degrading the overall image quality.
  arXiv  Detail & Related papers  (2023-07-12T07:48:29Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.