This study systematically uncovers and quantitatively evaluates the pervasive Orientalist biases in text-to-image (T2I) and text-to-video (T2V) generation models through a sociocultural lens grounded in postcolonial Orientalist theoretical frameworks. We identify systematic biases in the visual representations produced by multimodal generative models, including hyper-exoticization and temporal alienation. These biases mirror colonial-era narratives and undermine equitable sociocultural communication. Through empirical analysis of 8 mainstream T2I models and 4 T2V models, we demonstrate that culturally neutral prompts related to China consistently generate visual outputs embedded with Orientalist biases. We develop a novel visual question answering (VQA) framework as an evaluation metric, leveraging state-of-the-art vision-language model (VLM) to establish the first automated quantitative assessment methodology for such biases. A mitigation framework employing large language model (LLM) is proposed and experimentally validated. This interdisciplinary work illuminates the societal implications of multimodal generative models while advancing efforts toward fair and inclusive social computing.
论文检索
输入标题、作者或关键词，从 1,620 篇学术成果中精准定位
Generative AI models have substantially improved the realism of synthetic media, yet their misuse through sophisticated DeepFakes poses significant risks. Despite recent advances in deepfake detection, fairness remains inadequately addressed, enabling deepfake markers to exploit biases against specific populations. While previous studies have emphasized group-level fairness, individual fairness (i.e., ensuring similar predictions for similar individuals) remains largely unexplored. In this work, we identify for the first time that the original principle of individual fairness fundamentally fails in the context of deepfake detection, revealing a critical gap previously unexplored in the literature. To mitigate it, we propose the first generalizable framework that can be integrated into existing deepfake detectors to enhance individual fairness and generalization. Extensive experiments conducted on leading deepfake datasets demonstrate that our approach significantly improves individual fairness while maintaining robust detection performance, outperforming state-of-the-art methods. The code is available at: https://github.com/Purdue-M2/Individual-Fairness-Deepfake-Detection.
The rapid advancement of multimodal large language models (MLLMs) has led to breakthroughs in various applications, yet their security remains a critical challenge. One pressing issue involves unsafe image-query pairs-jailbreak inputs specifically designed to bypass security constraints and elicit unintended responses from MLLMs. Compared to general multimodal data, such unsafe inputs are relatively sparse, which limits the diversity and richness of training samples available for developing robust defense models. Meanwhile, existing guardrail-type methods rely on external modules to enforce security constraints but fail to address intrinsic vulnerabilities within MLLMs. Traditional supervised fine-tuning (SFT), on the other hand, often over-refuses harmless inputs, compromising general performance. Given these challenges, we propose Secure Tug-of-War (SecTOW), an innovative iterative defense-attack training method to enhance the security of MLLMs. SecTOW consists of two modules: a defender and an auxiliary attacker, both trained iteratively using reinforcement learning (GRPO). During the iterative process, the attacker identifies security vulnerabilities in the defense model and expands jailbreak data. The expanded data are then used to train the defender, enabling it to address identified security vulnerabilities. We also design reward mechanisms used for GRPO to simplify the use of response labels, reducing dependence on complex generative labels and enabling the efficient use of synthetic data. Additionally, a quality monitoring mechanism is used to mitigate the defender's over-refusal of harmless inputs and ensure the diversity of the jailbreak data generated by the attacker. Experimental results on safety-specific and general benchmarks demonstrate that SecTOW significantly improves security while preserving general performance. Warning: This paper contains offensive and unsafe content.
Democratic mediation serves as a vital mechanism for resolving social conflicts; however, current practices encounter three critical limitations: (1) inefficient operations, wherein traditional laborintensive mediation processes are both time-consuming and inefficient; (2) theoretical gaps, as prevailing mediation theories fail to explore the underlying causes of conflicts; and (3) inadequate analysis, with existing digital tools lacking comprehensive conflict mediation capabilities and primarily focusing on singular data types. To address these limitations, we introduce the Normative Social Simulator for Democratic Mediation, referred to as Norm Mediat. This framework is specifically designed to simulate democratic mediation, incorporating social norms. Central to this framework is the integration of normative reasoning into the mediation process, which enhances the ability to understand individuals' intrinsic needs and identify the root causes of conflicts. The framework comprises two essential components: (1) Dynamic Multimodal Conflict Modeling (DMCM), which generates the initial dataset of conflict interactions; and (2) Norm-Aware Iterative Mediation (NAIM), which implements an iterative democratic mediation process through norm awareness. The results of our human evaluation underscore the effectiveness of our norm-driven mediation strategies. This research significantly contributes to computational social science by providing a comprehensive methodological framework for simulating democratic processes and offering a benchmark dataset for conflict resolution studies.
The rapid advancement of generative AI has raised concerns about the authenticity of digital images, as highly realistic fake images can now be generated at low cost, potentially increasing societal risks. In response, several datasets have been established to train detection models aimed at distinguishing AI-generated images from real ones. However, existing datasets suffer from limited generalization, low image quality, overly simple prompts, and insufficient image diversity. To address these limitations, we propose a high-quality, large-scale dataset comprising over 730,000 images across multiple categories, including both real and AI-generated images. The generated images are synthesized via state-of-the-art methods, including text-to-image generation (guided by over 10,000 carefully designed prompts), image inpainting, image refinement, and face swapping. Each generated image is annotated with its generation method and category. Inpainting images further include binary masks to indicate inpainted regions, providing rich metadata for analysis. Compared to existing datasets, detection models trained on our dataset demonstrate superior generalization capabilities. Our dataset not only serves as a strong benchmark for evaluating detection methods but also contributes to advancing the robustness of AI-generated image detection techniques. Building upon this, we propose a lightweight detection method based on image noise entropy, which transforms the original image into an entropy tensor of Non-Local Means (NLM) noise before classification. Extensive experiments demonstrate that models trained on our dataset achieve strong generalization, and our method delivers competitive performance, establishing a solid baseline for future research. The dataset and source code are publicly available at https://real-hd.github.io.
Bitwise Vision AutoRegressive (BVAR) Model, as a distinguished source of young blood, has been taking the lead in the track of text-to-image synthesis, which at the same time raises legal and ethnic concerns such as copyright and authenticity. However, existing methods mainly focus on watermarking within diffusion models, which rely on the distinctive attributes of diffusion steps and cannot be directly transferred to new circumstances. To this end, we propose Safe-BVAR, the first watermark framework to embed bit strings during image generation in BVAR. Our study discovers the local similarity of the inferenced latent feature and the element-wise robustness of image autoencoder. Therefore, combined with the residual-accumulative nature of BVAR, we propose a novel Late Stage Residual Implanter to embed watermark and extract the information based on Local Contextual Extractor. Furthermore, we propose a Distributed Rotational Arranger to enhance watermark against local distortions. Our method is training-free and plug-and-play. Meanwhile, it can be easily applied to flexible-sized images. We evaluate the robustness and invisibility of the watermark, showing that it can resist common image attacks and cast inappreciable influence on the image.
Latent Diffusion Unlearning: Protecting Against Unauthorized Personalization Through Trajectory Shifted Perturbations
PDF ↗Text-to-image diffusion models have demonstrated remarkable effectiveness in rapid and high-fidelity personalization, even when provided with only a few user images. However, the effectiveness of personalization techniques has lead to concerns regarding data privacy, intellectual property protection, and unauthorized usage. To mitigate such unauthorized usage and model replication, the idea of generating ''unlearnable'' training samples utilizing image poisoning techniques has emerged. Existing methods for this have limited imperceptibility as they operate in the pixel space which results in images with noise and artifacts. In this work, we propose a novel model-based perturbation strategy that operates within the latent space of diffusion models. Our method alternates between denoising and inversion while modifying the starting point of the denoising trajectory: of diffusion models. This trajectory-shifted sampling ensures that the perturbed images maintain high visual fidelity to the original inputs while being resistant to inversion and personalization by downstream generative models. This approach integrates unlearnability into the framework of Latent Diffusion Models (LDMs), enabling a practical and imperceptible defense against unauthorized model adaptation. We validate our approach on four benchmark datasets to demonstrate robustness against state-of-the-art inversion attacks. Results demonstrate that our method achieves significant improvements in imperceptibility (~8% - 10% on perceptual metrics including PSNR, SSIM, and FID) and robustness (~10% on average across five adversarial settings), highlighting its effectiveness in safeguarding sensitive data. https://github.com/naresh-ub/unlearnable_samples.
Existing synthetic image detection approaches can be categorized into three paradigms: spatial, frequency, and fingerprint-based methods. Our analysis reveals a fundamental commonality across these paradigms: a significant reliance on high-frequency image components. This observation highlights the discriminative power of high-frequency information for this task and provides a strong rationale for learning generalized artifact representations based on multi-modal fusion strategies. Building on this insight, we introduce a multi-modal high-frequency interactive detection framework for general synthetic image detection. This framework explicitly integrates high-frequency information from both the spatial and frequency domains. Specifically, its spatial processing branch incorporates a novel high-frequency self-enhancement module to bolster local high-frequency representations. Concurrently, the frequency processing branch utilizes a multi-scale frequency information enhancement module to capture diverse contextual cues. At the feature fusion stage, we propose a pooling-guided cross-modal high-frequency interaction module, which dynamically weights cross-modal information to further reinforce salient high-frequency representations. Extensive experiments on public datasets demonstrate that our proposed framework achieves state-of-the-art performance in real-world detection scenarios.
Subspace-based DPSGD has emerged as a robust solution for alleviating excessive noise in high-dimensional, privacy-preserving visual learning. It achieves a favorable privacy-utility balance by projecting privatized gradients onto a low-rank subspace derived from in-distribution public data. However, relying solely on limited public data can narrow the diversity of the anchored subspace and induce over-memorization into model training, leading to suboptimal private optimization.To overcome these limitations, we propose a synthesis-augmented subspace-based DPSGD framework, SAS-DPSGD, which integrates synthetic data into the subspace construction process. We quantitatively analyze the private optimization performance using the subspace derived from the mixed public and synthetic data, revealing the benefits as well as the saturation effects of incorporating synthetic data in private visual learning. To the best of our knowledge, this is the first work to provide a unified theoretical guarantee to synthesis-augmented subspace-based DPSGD. Moreover, we design an early projection mechanism within our framework that projects the gradient onto the subspace before performing gradient clipping. This mechanism effectively reduces the gradient clipping bias and lowers the synthetic data requirement, resulting in a faster convergence rate. Extensive experiments on two real-world datasets validate that SAS-DPSGD outperforms nine baselines by up to 9.78% in accuracy and can reduce the amount of synthetic data required by 66.7%.
Assessing whether AI-generated images are substantially similar to copyrighted works is a crucial step in resolving copyright disputes. In this paper, we propose CopyJudge, an automated copyright infringement identification framework that leverages large vision-language models (LVLMs) to simulate practical court processes for determining substantial similarity between copyrighted images and those generated by text-to-image diffusion models. Specifically, we employ an abstraction-filtration-comparison test framework with multi-LVLM debate to assess the likelihood of infringement and provide detailed judgment rationales. Based on the judgments, we further introduce a general LVLM-based mitigation strategy that automatically optimizes infringing prompts by avoiding sensitive expressions while preserving the non-infringing content. Besides, our approach can be enhanced by exploring non-infringing noise vectors within the diffusion latent space via reinforcement learning, even without modifying the original prompts.Experimental results show that our identification method achieves comparable state-of-the-art performance, while offering superior generalization and interpretability across various forms of infringement, and that our mitigation method could more effectively mitigate memorization and IP infringement without losing non-infringing expressions.
With the surge of large language models (LLMs), Large Vision-Language Models (VLMs)-which integrate vision encoders with LLMs for accurate visual grounding-have shown great potential in tasks like generalist agents and robotic control. However, VLMs are typically trained on massive web-scraped images, raising concerns over copyright infringement and privacy violations, and making data auditing increasingly urgent. Membership inference (MI), which determines whether a sample was used in training, has emerged as a key auditing technique, with promising results on open-source VLMs like LLaVA (AUC > 80%). In this work, we revisit these advances and uncover a critical issue: current MI benchmarks suffer from distribution shifts between member and non-member images, introducing shortcut cues that inflate MI performance. We further analyze the nature of these shifts and propose a principled metric based on optimal transport to quantify the distribution discrepancy. To evaluate MI in realistic settings, we construct new benchmarks with i.i.d. member and non-member images. Existing MI methods fail under these unbiased conditions, performing only marginally better than chance. Further, we explore the theoretical upper bound of MI by probing the Bayes Optimality within the VLM's embedding space and find the irreducible error rate remains high. Despite this pessimistic outlook, we analyze why MI for VLMs is particularly challenging and identify three practical scenarios-fine-tuning, access to ground-truth texts, and set-based inference-where auditing becomes feasible. Our study presents a systematic view of the limits and opportunities of MI for VLMs, providing guidance for future efforts in trustworthy data auditing. Code and data will be available at https://github.com/GradOpt/Revisiting-VLM-MIA\faGithub.
The success of diffusion models in text-to-image (T2I) generation has made it urgent to remove unwanted concepts, such as copyrighted, offensive, and unsafe ones, from pre-trained models in an accurate, timely, and cost-effective manner. However, limited by the inherent optimization perspective, existing methods have two major problems. Firstly, they overlook maintaining the global visual style during the erasure process, leading to significant style shifts. Secondly, excessive concept erasure causes relevant content to disappear or generates substitutes unrelated to the original object's attributes. Compared to other methods, our proposed ICE has unique advantages, as it can generate diverse visual features and achieve a balance between concept erasure and maintaining the semantic content of the target object. This is mainly achieved through our well-designed non-erasable features protector (NEFP) and augmented invariant constraints (AIC). Specifically, we enhance the protection of feature information by embedding an augmented orthogonal anchor concept matrix. Meanwhile, under controlled constraints, we introduce invariants into the embedding space to retain key semantics. This work specifically emphasizes the importance of focusing on feature expression and semantic protection in the concept erasure task for fully unleashing the performance of T2I models.
The rapid advancement of deepfake technology has led to an increasing frequency of crisis incidents stemming from its misuse. However, existing forgery detection methods often suffer from poor cross-domain generalization due to overfitting to specific forgery cues inherent in their training datasets. Through an in-depth analysis, we identify that low-loss, overfitted features hinder models from capturing broadly applicable patterns necessary for effective generalization. To overcome this limitation, we introduce Knowledge Negative Distillation (KND), a simple yet powerful teacher-student framework, designed to encourage the student model to acquire knowledge beyond the teacher's existing scope. Specifically, we guide the student model to avoid the teacher's overfitted features by maximizing a cross-entropy loss computed from the teacher's probability distributions during the training for the target task. Additionally, we propose an adaptive fusion mechanism that integrates the extensible student features with the teacher's features, weighted and guided by their respective probability distributions. Extensive experimental results validate the superior performance of KND, demonstrating state-of-the-art capabilities across multiple benchmarks. Moreover, the extensibility and universality of KND underscore its potential applicability to a broader range of cross-domain problems characterized by significant overfitting challenges.
The detection and grounding of manipulated content in multimodal data has emerged as a critical challenge in media forensics. While existing benchmarks demonstrate technical progress, they suffer from misalignment artifacts that poorly reflect real-world manipulation patterns: practical attacks typically maintain semantic consistency across modalities, whereas current datasets artificially disrupt cross-modal alignment, creating easily detectable anomalies. To bridge this gap, we pioneer the detection of semantically-coordinated manipulations where visual edits are systematically paired with semantically consistent textual descriptions. Our approach begins with constructing the first Semantic-Aligned Multimodal Manipulation (SAMM) dataset, generated through a two-stage pipeline: 1) applying state-of-the-art image manipulations, followed by 2) generation of contextually-plausible textual narratives that reinforce the visual deception. Building on this foundation, we propose a Retrieval-Augmented Manipulation Detection and Grounding (RamDG) framework. RamDG commences by harnessing external knowledge repositories to retrieve contextual evidence, which serves as the auxiliary texts and encoded together with the inputs through our image forgery grounding and deep manipulation detection modules to trace all manipulations. Extensive experiments demonstrate our framework significantly outperforms existing methods, achieving 2.06% higher detection accuracy on SAMM compared to state-of-the-art approaches. The dataset and code are publicly available at https://github.com/shen8424/SAMM-RamDG-CAP.
Residential design is a complex and open-ended problem that requires designers to integrate diverse types of input information while adhering to stringent energy consumption standards. However, most current research in this field focuses on generating floor plans from a limited set of input types, often neglecting to incorporate energy-related physical constraints. Existing approaches are limited by: (1) the lack of multimodal datasets in this domain, (2) the absence of comprehensive residential energy consumption data, and (3) the challenges associated with effectively integrating multiple input types into a unified model. To address these challenges, we propose MRED-14, the first large-scale Multimodal Residential Energy Dataset, comprising 14 input types, including energy consumption values, vector drawings, and textual descriptions, paired with 41,280 high-quality residential floor plans that have been scored and annotated by human experts. Based on this dataset, we introduce the LER-net model, which can flexibly adapt to various input types and generate low-energy residential floor plans. Experimental results demonstrate that LER-net outperforms existing models, achieving state-of-the-art performance under the same input conditions. In addition, the energy consumption of the generated floor plans is reduced by 5.1% compared to the actual residential designs. Further expert evaluations confirm the LER-net model's feasibility for use in residential design.
Federated face generation technology leverages decentralized private data to achieve high-quality face synthesis. However, regulations such as the GDPR confer users the right to be forgotten, necessitating the removal of contributions from specific clients in the global model. Existing generation model unlearning methods are primarily designed for centralized environments and are inadequate for addressing the constraints of data privacy storage and limited client computational resources in federated settings. To address this gap, we propose F2GU, the first federated unlearning framework specifically tailored for face generation models, enabling the effective removal of contributions associated with specific clients (identities) while ensuring privacy. Our proposed Generation Trajectory Redirection method dynamically guides the generation trajectory away from target identities, thereby effectively eliminating contributions from specific clients. Additionally, we devise a Mirroring-guided Trajectory Optimization strategy that constructs a mirror projection utilizing the retained client trajectory origins to ensure the generative capabilities of the model are preserved post-unlearning. We conduct extensive experiments on two mainstream face generation models (GAN and Diffusion Model) across three different datasets.The results indicate that our method demonstrates superior performance in both the success rate of identity unlearning and the preservation of generation quality. The code can be available at https://github.com/FanQi-AI/FFGU.
Various (text) prompt filters and (image) safety checkers have been implemented to mitigate the misuse of Text-to-Image (T2I) models in creating Not-Safe-For-Work (NSFW) content. In order to expose potential security vulnerabilities of such safeguards, multimodal jailbreaks have been studied. However, existing jailbreaks are limited to prompt-specific and image-specific perturbations, which suffer from poor scalability and time-consuming optimization. To address these limitations, we propose Universally Unfiltered and Unseen (U3)-Attack, a multimodal jailbreak attack method against T2I safeguards. Specifically, U3-Attack optimizes an adversarial patch on the image background to universally bypass safety checkers and optimizes a safe paraphrase set from a sensitive word to universally bypass prompt filters while eliminating redundant computations. Extensive experimental results demonstrate the superiority of our U3-Attack on both open-source and commercial T2I models. For example, on the commercial Runway-inpainting model with both prompt filter and safety checker, our U3-Attack achieves approximately 4× higher success rates than the state-of-the-art multimodal jailbreak attack, MMA-Diffusion. Content Warning: This paper includes examples of NSFW content.
While the widespread adoption of diffusion models in image generation has showcased remarkable capabilities, it has also inadvertently opened the door to malicious exploitation. Recent research has primarily concentrated on protecting images from the misuse of diffusion-based customized generation (CG). However, these approaches often overlook that image details can still be enhanced through diffusion-based super-resolution (SR) techniques, significantly increasing the risks of personal image leakage and abuse. To combat these multifaceted risks, we propose the Zero Matrix-guided Adaptive Image Vaccine (ZMAIV) framework. Specifically, we introduce the Self-attention Removal strategy, tailored for CG, which disrupts the model's core mechanism of focusing on sensitive spaces. Concurrently, the High-frequency Removal strategy is proposed to impede the high-frequency details reconstruction of SR. These defense strategies effectively dismantle the underlying mechanisms that facilitate unauthorized data extrapolation. Moreover, the proposed Adaptive Space Search Attack precisely targets critical spaces within images for vaccine injection, optimizing perturbation placement to minimize perturbation conflict while maintaining defense performance. Extensive experiments demonstrate that the proposed ZMAIV outperforms the state-of-the-arts in the aspects of simultaneously defending against diffusion-based CG and SR, affirming its superiority in safeguarding visual content against these dual threats.
Anomaly inspection plays a vital role in industrial manufacturing, but the scarcity of anomaly samples significantly limits the effectiveness of existing methods in tasks such as localization and classification. While several anomaly synthesis approaches have been introduced for data augmentation, they often struggle with low realism, inaccurate mask alignment, and poor generalization. To overcome these limitations, we propose Generate Aligned Anomaly (GAA), a region-guided, few-shot anomaly image-mask pair generation framework. GAA leverages the strong priors of a pretrained latent diffusion model to generate realistic, diverse, and semantically aligned anomalies using only a small number of samples. The framework first employs Localized Concept Decomposition to jointly model the semantic features and spatial information of anomalies, enabling flexible control over the type and location of anomalies. It then utilizes Adaptive Multi-Round Anomaly Clustering to perform fine-grained semantic clustering of anomaly concepts, thereby enhancing the consistency of anomaly representations. Subsequently, a region-guided mask generation strategy ensures precise alignment between anomalies and their corresponding masks, while a low-quality sample filtering module is introduced to further improve the overall quality of the generated samples. Extensive experiments on the MVTec AD and LOCO datasets demonstrate that GAA achieves superior performance in both anomaly synthesis quality and downstream tasks such as localization and classification.
Social biases in text-to-image models have drawn increasing attention, yet existing debiasing efforts often focus solely on either the textual (e.g., CLIP) or visual (e.g., U-Net) space. This unimodal perspective introduces two major challenges: (i) Debiasing only the textual space fails to control visual outputs, often leading to pseudo- or over-corrections due to unaddressed visual biases during denoising; (ii) Debiasing only the visual space can cause modality conflicts when biases in textual and vision are misaligned, degrading the quality and consistency of generated images. To address these issues, we propose a Bimodal ADaptive Guidance DEbiasing within Textual and Visual Spaces (BADGE). First, BADGE quantifies attribute-level bias inclination in both modalities, providing precise guidance for subsequent mitigation. Second, to avoid pseudo/over-correction and modality conflicts, the quantified bias degree is used as the debiasing strength for adaptive guidance, enabling fine-grained correction tailored to discrete attribute concepts.Extensive experiments demonstrate that BADGE significantly enhances fairness across intra- and inter-category attributes (e.g., gender, skin tone, age, and their interaction) while preserving high image fidelity. *Our project page is at https://badgediffusion.github.io/