Text-to-Image (T2I) models typically deploy safety mechanisms to prevent the generation of sensitive images. Unfortunately, recent jailbreaking attack methods manually design instructions for the LLM to generate adversarial prompts, which effectively exposing safety vulnerabilities of T2I models. However, existing methods have two limitations: 1) relying on manually exhaustive strategies for designing adversarial prompts, lacking a unified framework, and 2) requiring numerous queries to achieve a successful attack, limiting their practical applicability. To address this issue, we propose Reason2Attack~(R2A), which aims to enhance the effectiveness and efficiency of the LLM in jailbreaking attacks. Specifically, we first use Frame Semantics theory to systematize existing manually crafted strategies and propose a unified generation framework to generate CoT adversarial prompts step by step. Following this, we propose a two-stage LLM reasoning training framework guided by the attack process. In the first stage, the LLM is fine-tuned with CoT examples generated by the unified generation framework to internalize the adversarial prompt generation process grounded in Frame Semantics. In the second stage, we incorporate the jailbreaking task into the LLM's reinforcement learning process, guided by the proposed attack process reward function that balances prompt stealthiness, effectiveness, and length, enabling the LLM to understand T2I models and safety mechanisms. Extensive experiments on various T2I models with safety mechanisms, and commercial T2I models, show the superiority and practicality of R2A.

Using risky text prompts, such as pornography and violent prompts, to test the safety of text-to-image (T2I) models is a critical task. However, existing risky prompt datasets are limited in three key areas: 1) limited risky categories, 2) coarse-grained annotation, and 3) low effectiveness. To address these limitations, we introduce T2I-RiskyPrompt, a comprehensive benchmark designed for evaluating safety-related tasks in T2I models. Specifically, we first develop a hierarchical risk taxonomy, which consists of 6 primary categories and 14 fine-grained subcategories. Building upon this taxonomy, we construct a pipeline to collect and annotate risky prompts. Finally, we obtain 6,432 effective risky prompts, where each prompt is annotated with both hierarchical category labels and detailed risk reasons. Moreover, to facilitate the evaluation, we propose a reason-driven risky image detection method that explicitly aligns the MLLM with safety annotations. Based on T2I-RiskyPrompt, we conduct a comprehensive evaluation of eight T2I models, nine defense methods, five safety filters, and five attack strategies, offering nine key insights into the strengths and limitations of T2I model safety. Finally, we discuss potential applications of T2I-RiskyPrompt across various research fields.

Text-to-image (T2I) models aim to output high-quality images based on the input prompt. It commonly incorporates safety filters to prevent the generation of sensitive images. In this study, we aim to reveal the safety vulnerabilities of black-box T2I models in a real-world scenario. To address the challenges of bypassing filters and mitigating the trade-off between stealthiness and effectiveness, we investigate the concepts from Semiotics and apply cultural elements in the design of adversarial prompts. Specifically, we propose CAAM, a Culture-based Adversarial Attack on T2I Models with a generation module, which collaborates with multiple LLM-based agents to generate fluent adversarial prompts, and a feedback module, which provides feedback to compromise the stealthiness and effectiveness. Extensive experiments with open-source and commercial T2I models are conducted to demonstrate the effectiveness of our method. This paper includes model-generated content that may contain offensive or distressing material.

Although there are several food-level benchmarks for food-related learning, the lack of fine-grained ingredient annotation significantly impedes progress in food scene understanding. In this study, we focus on Chinese food understanding which involves fine-grained ingredient detection and cross-modal ingredient retrieval. Specifically, to support studies on Chinese food understanding, we build the first cross-modal ingredientlevel dataset called CMIngre, which contains 8,001 image-text pairs from three different sources, i.e. dishes, recipes, and usergenerated content, covering 429 distinct ingredients and 95,290 bounding boxes. Based on CMIngre, we evaluate the performance of traditional CNN-based detection algorithms and transformerbased pre-trained large models for ingredient detection. We also propose baseline methods for the cross-modal ingredient retrieval task in both the end-to-end and two-stage settings. Extensive experiments on CMIngre demonstrate the effectiveness of our proposed methods on food understanding

Unsupervised cross-domain 3D model retrieval aims to retrieve unlabeled 3D models (target domain) using labeled 2D images (source domain). Domain adaptation approaches have shown impressive performance for cross-domain 3D model retrieval. However, conventional methods typically represent samples from different domains as deterministic points, overlooking the diversity in sample characteristics and relationships. These approaches lead to challenges in achieving a robust representation of both samples and categories. To address above challenges, we propose a dual-stage uncertainty modeling (DSUM) for unsupervised cross-domain 3D model retrieval, which utilizes Gaussian distribution to effectively model the uncertainty characteristics in both sample and class and obtain the robust and domain-invariant representations. Specifically, in the multi-view uncertainty encoding stage, we discard the conventional pooling operations and utilize the uncertainty modeling among multiple views to fuse the common and specific information of 2D images and 3D models. In the cross-domain feature alignment stage, we adopt the Gaussian distribution of samples belonging to the same category, which can well maintain the sample diversity as well as facilitate to eliminate the domain discrepancy. Our method achieves improvements of 2.61% and 2.65% in terms of FT on two cross-domain datasets, respectively, verifying its superiority through extensive qualitative and quantitative experiments.

Unsupervised 2D image-based 3D shape retrieval aims to match the similar 3D unlabeled shapes when given a 2D labeled sample. Although a lot of methods have made a certain degree of progress, the performance of this task is still restricted due to the lack of target labels resulting in tremendous domain gap. In this paper, we aim to explore the discriminative representation of the unlabeled target 3D shapes and facilitate the procedure of domain adaptation by taking full advantage of multi-view information. To achieve the above goals, we propose an effective self-supervised auxiliary domain alignment (SADA) for unsupervised 2D image-based 3D shape retrieval. SADA mainly contains multi-view guided self-supervised feature learning and two auxiliary domain alignments, including intermediate domain alignment and multi-domain alignment. Firstly, we group multiple views of each 3D shape into two sub-target domains based on the view similarities and regard each other as the constraint to optimize the feature learning in an unsupervised manner. To reduce the difficulty of directly aligning the domain discrepancy, we combine the source labeled samples and target samples (pseudo labels) with the same category to generate an intermediate domain, which translates the source-target alignment into source-intermediate and intermediate-target alignments. Moreover, to explore the inner characteristics of target 3D shapes and provide more clues for better adaptation, multi-domain alignment is proposed to convert the source and single target domain alignment to the source and multiple target domain (one target domain and two sub-target domains) alignments. The adversarial training and semantic alignment are employed to fully excavate the relations between source domain and multiple target domains. Experiments on two challenging datasets show that the proposed method achieves competing performance in the unsupervised 2D image-based 3D shape retrieval task.