论文检索

Lu Zhang, Liang Zeng

The rapid proliferation of AI-generated images necessitates effective watermarking techniques to protect intellectual property and detect fraudulent content. While existing training-based watermarking methods show promise, they often struggle with generalization across diverse prompts, introduce visible artifacts, and require substantial external data for retraining on new model variants. To this end, we propose Modular Self-Augmented Training for Latent Diffusion Models (MSAT-LDM), a novel and transferable watermarking framework. MSAT-LDM integrates two key components: (1) Self-Augmented Training (SAT) leverages an internally generated "free generation" distribution to train the watermark module, aligning the training and testing phases without relying on external data. We theoretically demonstrate that this design improves generalization by inducing a tighter generalization bound. (2) Modular watermark architecture is a plug-and-play module that can be independently fine-tuned, enabling efficient adaptation to various fine-tuned backbones or LoRA-enhanced variants with minimal overhead. Extensive experiments show that MSAT-LDM achieves robust watermarking, significantly improves the quality of watermarked images across diverse prompts, and exhibits strong transfer performance--all without the need for external training data.

Chenyu Zhang, Tairen Zhang, Lanjun Wang, Ruidong Chen, Wenhui Li, Anan Liu

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.

Chenyu Zhang, Lanjun Wang, Yiwen Ma, Wenhui Li, Guoqing Jin, Anan Liu

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.

Liu Yu, Zhonghao Chen, Ping Kuang, Zhikun Feng, Fan Zhou, Lan Wang, Gillian Dobbie

Object hallucination remains a critical challenge in Large Vision-Language Models (LVLMs), where models generate content inconsistent with visual inputs. Existing language-decoder based mitigation approaches often regulate visual or textual attention independently, overlooking their interaction as two key causal factors. To address this, we propose Owl (Bi-mOdal attention reWeighting for Layer-wise hallucination mitigation), a causally-grounded framework that models hallucination process via a structural causal graph, treating decomposed visual and textual attentions as mediators. We introduce VTACR (Visual-to-Textual Attention Contribution Ratio), a novel metric that quantifies the modality contribution imbalance during decoding. Our analysis reveals that hallucinations frequently occur in low-VTACR scenarios, where textual priors dominate and visual grounding is weakened. To mitigate this, we design a fine-grained attention intervention mechanism that dynamically adjusts token- and layer-wise attention guided by VTACR signals. Finally, we propose a dual-path contrastive decoding strategy: one path emphasizes visually grounded predictions, while the other amplifies hallucinated ones -- letting visual truth shine and hallucination collapse. Experimental results on the POPE and CHAIR benchmarks show that Owl achieves significant hallucination reduction, setting a new SOTA in faithfulness while preserving vision-language understanding capability. Our code is available at https://github.com/CikZ2023/OWL

Adeel Yousaf, Joseph Fioresi, James Beetham, Amrit Singh Bedi, Mubarak Shah

Improving the safety of vision-language models like CLIP via fine-tuning often comes at a steep price, causing significant drops in their generalization performance. We find this trade-off stems from rigid alignment strategies that force unsafe concepts toward single, predefined safe targets, disrupting the model's learned semantic structure. To address this, we propose a proximity-aware approach: redirecting unsafe concepts to their semantically closest safe alternatives to minimize representational change. We introduce SafeR-CLIP, a fine-tuning framework that applies this principle of minimal intervention. SafeR-CLIP successfully reconciles safety and performance, recovering up to 8.0% in zero-shot accuracy over prior methods while maintaining robust safety. To support more rigorous evaluation, we also contribute NSFWCaps, a new benchmark of 1,000 highly-aligned pairs for testing safety under distributional shift. Our work shows that respecting the geometry of pretrained representations is key to achieving safety without sacrificing performance.

Xi Ye, Yiwen Liu, Lina Wang, Run Wang, Geying Yang, Yufei Hou, Jiayi Yu

Text-to-image (T2I) models have raised increasing safety concerns due to their capacity to generate NSFW and other banned objects. To mitigate these risks, safety filters and concept removal techniques have been introduced to block inappropriate prompts or erase sensitive concepts from the models. However, all the existing defense methods are not well prepared to handle diverse adversarial prompts. In this work, we introduce MacPrompt, a novel black-box and cross-lingual attack that reveals previously overlooked vulnerabilities in T2I safety mechanisms. Unlike existing attacks that rely on synonym substitution or prompt obfuscation, MacPrompt constructs macaronic adversarial prompts by performing cross-lingual character-level recombination of harmful terms, enabling fine-grained control over both semantics and appearance. By leveraging this design, MacPrompt crafts prompts with high semantic similarity to the original harmful inputs (up to 0.96) while bypassing major safety filters (up to 100%). More critically, it achieves attack success rates as high as 92% for sex-related content and 90\% for violence, effectively breaking even state-of-the-art concept removal defenses. These results underscore the pressing need to reassess the robustness of existing T2I safety mechanisms against linguistically diverse and fine-grained adversarial strategies. Warning: This paper includes sensitive examples (e.g., adult, violent, or illegal content). Unsafe images are masked but may still be disturbing.

Yuan Yao, Jin Song, Jian Jin

As valuable digital assets, deep neural networks necessitate robust ownership protection, positioning neural network watermarking (NNW) as a promising solution. Among various NNW approaches, weight-based methods are favored for their simplicity and practicality; however, they remain generally vulnerable to forging and overwriting attacks. To address those challenges, we propose *NeuralMark*, a robust method built around a *hashed watermark filter*. Specifically, we utilize a hash function to generate an irreversible binary watermark from a secret key, which is then used as a filter to select the model parameters for embedding. This design cleverly intertwines the embedding parameters with the hashed watermark, providing a robust defense against both forging and overwriting attacks. Average pooling is also incorporated to resist fine-tuning and pruning attacks. Furthermore, it can be seamlessly integrated into various neural network architectures, ensuring broad applicability. We theoretically analyze its security boundary and highlight the necessity of using a hashed watermark as a filtering mechanism. Empirically, we demonstrate its effectiveness and robustness across 13 distinct Convolutional and Transformer architectures, covering five image classification tasks and one text generation task.

Yiran Yang, Zhaowei Liu, Yuan Yuan, Yukun Song, Xiong Ma, Yinghao Song, Xiangji Zeng, Lu Sun, Yulu Wang, Hai Zhou 等

Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle–speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.

Yuan Xun, Xiaojun Jia, Xinwei Liu, Simeng Qin, Hua Zhang

Multimodal large reasoning models (MLRMs) have advanced visual-textual integration, enabling sophisticated human-AI interaction. While prior work has exposed MLRMs to visual jailbreaks, it remains underexplored how their reasoning capabilities reshape the security landscape under adversarial inputs. To fill this gap, we conduct a systematic security assessment of MLRMs and uncover a security-reasoning paradox: although deeper reasoning boosts cross‑modal risk recognition, it also creates cognitive blind spots that adversaries can exploit. We observe that MLRMs oriented toward human-centric service are highly susceptible to users' emotional cues during the deep-thinking stage, often overriding safety protocols or built‑in safety checks under high emotional intensity. Inspired by this key insight, we propose EmoAgent, an autonomous adversarial emotion-agent that orchestrates exaggerated affective prompts to hijack reasoning pathways. Even when visual risks are correctly identified, models can still produce harmful completions through emotional misalignment. We further identify persistent high-risk failure modes in transparent deep-thinking scenarios, such as MLRMs generating harmful reasoning masked behind seemingly safe responses. These failures expose misalignments between internal inference and surface-level behavior, eluding existing content-based safeguards. To quantify these risks, we introduce three metrics: (1) Risk-Reasoning Stealth Score (RRSS) for harmful reasoning beneath benign outputs; (2) Risk-Visual Neglect Rate (RVNR) for unsafe completions despite visual risk recognition; and (3) Refusal Attitude Inconsistency (RAIC) for evaluating refusal unstability under prompt variants. Extensive experiments on advanced MLRMs demonstrate the effectiveness of EmoAgent and reveal deeper emotional cognitive misalignments in model safety.

Jing Xue, Zhishen Sun, Haishan Ye, Luo Luo, Xiangyu Chang, Guang Dai

Membership inference attack (MIA) has become one of the most widely used and effective methods for evaluating the privacy risks of machine learning models. This attack aims to determine whether a specific sample is part of the model's training set by analyzing the model's output. While traditional membership inference attacks focus on leveraging the model’s posterior output, such as confidence on the target sample, we propose IMIA, a novel attack strategy that utilizes the process of generating adversarial samples to infer membership. We propose to infer the member properties of the target sample using the number of iterations required to generate its adversarial sample. We conduct experiments across multiple models and datasets, and our results demonstrate that the number of iterations for generating an adversarial sample is a reliable feature for membership inference, achieving strong performance both in black-box and white-box attack scenarios. This work provides a new perspective for evaluating model privacy and highlights the potential of adversarial example-based features for privacy leakage assessment.

Siyuan Xu, Yibing Liu, Peilin Chen, Yung-Hui Li, Shiqi Wang, Sam Kwong

Privacy leakage in Multimodal Large Language Models (MLLMs) has long been an intractable problem. Existing studies, though effectively obscure private information in MLLMs, often overlook the evaluation of authenticity and recovery quality of user privacy. To this end, this work uniquely focuses on the critical challenge of how to restore surrogate-driven protected data in diverse MLLM scenarios. We first bridge this research gap by contributing the SPPE (Surrogate Privacy Protected Editable) dataset, which includes a wide range of privacy categories and user instructions to simulate real MLLM applications. This dataset offers protected surrogates alongside their various MLLM-edited versions, thus enabling the direct assessment of privacy recovery quality. By formulating privacy recovery as a guided generation task conditioned on complementary multimodal signals, we further introduce a unified approach that reliably reconstructs private content while preserving the fidelity of MLLM-generated edits. The experiments on both SPPE and InstructPix2Pix further show that our approach generalizes well across diverse visual content and editing tasks, achieving a strong balance between privacy protection and MLLM usability.

Naen Xu, Jinghuai Zhang, Changjiang Li, Hengyu An, Chunyi Zhou, Jun Wang, Boyu Xu, Yuyuan Li, Tianyu Du, Shouling Ji

Large vision-language models (LVLMs) have achieved remarkable advancements in multimodal reasoning tasks. However, their widespread accessibility raises critical concerns about potential copyright infringement. Will LVLMs accurately recognize and comply with copyright regulations when encountering copyrighted content (i.e., user input, retrieved documents) in the context? Failure to comply with copyright regulations may lead to serious legal and ethical consequences, particularly when LVLMs generate responses based on copyrighted materials (e.g., retrieved book experts, news reports). In this paper, we present a comprehensive evaluation of various LVLMs, examining how they handle copyrighted content – such as book excerpts, news articles, music lyrics, and code documentation when they are presented as visual inputs. To systematically measure copyright compliance, we introduce a large-scale benchmark dataset comprising 50,000 multimodal query-content pairs designed to evaluate how effectively LVLMs handle queries that could lead to copyright infringement. Given that real-world copyrighted content may or may not include a copyright notice, the dataset includes query-content pairs in two distinct scenarios: with and without a copyright notice. For the former, we extensively cover four types of copyright notices to account for different cases. Our evaluation reveals that even state-of-the-art closed-source LVLMs exhibit significant deficiencies in recognizing and respecting the copyrighted content, even when presented with the copyright notice. To solve this limitation, we introduce a novel tool-augmented defense framework for copyright compliance, which reduces infringement risks in all scenarios. Our findings underscore the importance of developing copyright-aware LVLMs to ensure the responsible and lawful use of copyrighted content.

Zixun Xiong, Gaoyi Wu, Qingyang Yu, Mingyu Derek Ma, Lingfeng Yao, Miao Pan, Xiaojiang Du, Hao Wang

Given the high cost of large language model (LLM) training from scratch, safeguarding LLM intellectual property (IP) becomes increasingly crucial. As the standard paradigm for IP ownership verification, LLM fingerprinting thus plays a vital role in addressing this challenge. Existing LLM fingerprinting methods verify ownership by extracting or injecting model-specific features. However, they overlook potential attacks during the verification process, leaving them ineffective when the model thief fully controls the LLM's inference process. In such settings, attackers may share prompt-response pairs to enable fingerprint unlearning, or manipulate outputs to evade exact-match verification. We propose iSeal, the first fingerprinting method designed for reliable verification when the model thief controls the suspected LLM in an end-to-end manner. It injects unique features into both the model and an external module, reinforced by an error-correction mechanism and a similarity-based verification strategy. These components are resistant to verification-time attacks, including collusion-based fingerprint unlearning and response manipulation, backed by both theoretical analysis and empirical results. iSeal achieves 100% Fingerprint Success Rate (FSR) on 12 LLMs against more than 10 attacks, while baselines fail under unlearning and response manipulations.

Yuhan Xie, Chen Lyu

Split Federated Learning (SFL) is an emerging paradigm for privacy-preserving distributed learning. However, it remains vulnerable to sophisticated data poisoning attacks targeting local features, labels, smashed data, and model weights. Existing defenses, primarily adapted from traditional Federated Learning (FL), are less effective under SFL due to limited access to complete model updates. This paper presents HealSplit, the first unified defense framework tailored for SFL, offering end-to-end detection and recovery against five sophisticated types of poisoning attacks. HealSplit comprises three key components: (1) a topology-aware detection module that constructs graphs over smashed data to identify poisoned samples via topological anomaly scoring (TAS); (2) a generative recovery pipeline that synthesizes semantically consistent substitutes for detected anomalies, validated by a consistency validation student; and (3) an adversarial multi-teacher distillation framework trains the student using semantic supervision from a Vanilla Teacher and anomaly-aware signals from an Anomaly-Influence Debiasing (AD) Teacher, guided by the alignment between topological and gradient-based interaction matrices. Extensive experiments on four benchmark datasets demonstrate that HealSplit consistently outperforms ten state-of-the-art defenses, achieving superior robustness and defense effectiveness across diverse attack scenarios.

Yuankun Xie, Ruibo Fu, Xiaopeng Wang, Zhiyong Wang, Songjun Cao, Long Ma, Haonan Cheng, Long Ye

The rapid advancement of audio generation technologies has escalated the risks of malicious deepfake audio across speech, sound, singing voice, and music, threatening multimedia security and trust. While existing countermeasures (CMs) perform well in single-type audio deepfake detection (ADD), their performance declines in cross-type scenarios. This paper is dedicated to studying the all-type ADD task. We are the first to comprehensively establish an all-type ADD benchmark to evaluate current CMs, incorporating cross-type deepfake detection across speech, sound, singing voice, and music. Then, we introduce the prompt tuning self-supervised learning (PT-SSL) training paradigm, which optimizes SSL front-end by learning specialized prompt tokens for ADD, requiring 458× fewer trainable parameters than fine-tuning (FT). Considering the auditory perception of different audio types, we propose the wavelet prompt tuning (WPT)-SSL method to capture type-invariant auditory deepfake information from the frequency domain without requiring additional training parameters, thereby enhancing performance over FT in the all-type ADD task. To achieve an universally CM, we utilize all types of deepfake audio for co-training. Experimental results demonstrate that WPT-XLSR-AASIST achieved the best performance, with an average EER of 3.58% across all evaluation sets.

Huiyuan Xie, Chenyang Li, Huining Zhu, Chubin Zhang, Yuxiao Ye, Zhenghao Liu, Zhiyuan Liu

Legal reasoning is a fundamental component of legal analysis and decision-making. Existing computational approaches to legal reasoning predominantly rely on generic reasoning frameworks such as syllogism, which do not comprehensively examine the nuanced process of legal reasoning. Moreover, current research has largely focused on criminal cases, with insufficient modeling for civil cases. In this work, we present a novel framework to explicitly model legal reasoning in the analysis of Chinese tort-related civil cases. We first operationalize the legal reasoning process in tort analysis into the three-module LexChain framework, with each module consisting of multiple finer-grained sub-steps. Informed by the LexChain framework, we introduce the task of tort legal reasoning and construct an evaluation benchmark to systematically assess the critical steps within analytical reasoning chains for tort analysis. Leveraging this benchmark, we evaluate existing large language models for their legal reasoning ability in civil tort contexts. Our results indicate that current models still fall short in accurately handling crucial elements of tort legal reasoning. Furthermore, we introduce several baseline approaches that explicitly incorporate LexChain-style reasoning through prompting or post-training. The proposed baselines achieve significant improvements in tort-related legal reasoning and generalize well to related legal analysis tasks, demonstrating the value of explicitly modeling legal reasoning chains to enhance the reasoning capabilities of language models.

Yaxin Xiao, Qingqing Ye, Zi Liang, Haoyang Li, RongHua Li, Huadi Zheng, Haibo Hu

Machine learning models constitute valuable intellectual property, yet remain vulnerable to model extraction attacks (MEA), where adversaries replicate their functionality through black-box queries. Model watermarking counters MEAs by embedding forensic markers for ownership verification. Current black-box watermarks prioritize MEA survival through representation entanglement, yet inadequately explore resilience against sequential MEAs and removal attacks. Our study reveals that this risk is underestimated because existing removal methods are weakened by entanglement. To address this gap, we propose Watermark Removal attacK (WRK), which circumvents entanglement constraints by exploiting decision boundaries shaped by prevailing sample-level watermark artifacts. WRK effectively reduces watermark success rates by ≥88.79% across existing watermarking benchmarks. For robust protection, we propose Class-Feature Watermarks (CFW), which improve resilience by leveraging class-level artifacts. CFW constructs a synthetic class using out-of-domain samples, eliminating vulnerable decision boundaries between original domain samples and their artifact-modified counterparts (watermark samples). CFW concurrently optimizes both MEA transferability and post-MEA stability. Experiments across multiple domains show that CFW consistently outperforms prior methods in resilience, maintaining a watermark success rate of ≥70.15% in extracted models even under the combined MEA and WRK distortion, while preserving the utility of protected models.

Jie Xiao, Yuhao Huang, Yanjiao Gao, Aizhu Liu, Zhezhao Yang, Xinyue Yu, Qianwei Zhou, Fan Terry Zhang

Backdoor attacks on deep neural networks (DNNs) have garnered significant attention, particularly in edge computing applications. Given the complexity and opacity of DNNs, defending against backdoor attacks remains a formidable challenge. To address this, we propose CL-Guard, a dual-network-based defense framework designed to effectively eliminate potential backdoors in models. First, it leverages an inter-layer backpropagation algorithm to quantify each neuron's contribution to model prediction. Next, it constructs a critical neuron set through a recursive hierarchical partitioning method and an adaptive search strategy, identifying neurons critical to model prediction while minimizing the inclusion of backdoor-related neurons. Then, we perform sparse training on the non-critical neuron set, effectively strengthening the weights of critical neurons while disrupting the association between trigger features and backdoor-related neurons. Finally, we design a dual-network architecture that incorporates a fine-grained gradient backpropagation mechanism and dynamic collaborative learning, enabling the model to retain its original accuracy while preventing backdoor reactivation. The experimental results indicate that CL-Guard achieves an average Security Effectiveness Index (SEI) of approximately 95.42%, representing a 21.23% improvement over the state-of-the-art FT-SAM method.

Lingyun Xiang, Chengfu Ou, Xu He, Zhongliang Yang, Yuling Liu

Existing linguistic steganography methods primarily rely on content transformations to conceal secret messages. However, they often cause subtle yet looking-innocent deviations between normal and stego texts, posing potential security risks in real-world applications. To address this challenge, we propose a content-preserving linguistic steganography paradigm for perfectly secure covert communication without modifying the cover text. Based on this paradigm, we introduce CLstega (Content-preserving Linguistic steganography), a novel method that embeds secret messages through controllable distribution transformation. CLstega first applies an augmented masking strategy to locate and mask embedding positions, where MLM (masked language model)-predicted probability distributions are easily adjustable for transformation. Subsequently, a dynamic distribution steganographic coding strategy is designed to encode secret messages by deriving target distributions from the original probability distributions. To achieve this transformation, CLstega elaborately selects target words for embedding positions as labels to construct a masked sentence dataset, which is used to fine-tune the original MLM, producing a target MLM capable of directly extracting secret messages from the cover text. This approach ensures perfect security of secret messages while fully preserving the integrity of the original cover text. Experimental results demonstrate that CLstega can achieve a 100% extraction success rate, and outperforms existing methods in security, effectively balancing embedding capacity and security.

Zhengxian Wu, Juan Wen, Wanli Peng, Yinghan Zhou, Changtong Dou, Yiming Xue

Although existing backdoor defenses have gained success in mitigating backdoor attacks, they still face substantial challenges. In particular, most of them rely on large amounts of clean data to weaken the backdoor mapping but generally struggle with residual trigger effects, resulting in persistently high attack success rates (ASR). Therefore, in this paper, we propose a novel Backdoor defense method based on Directional mapping module and adversarial Knowledge Distillation (BeDKD), which balances the trade-off between defense effectiveness and model performance using a small amount of clean and poisoned data. We first introduce a directional mapping module to identify poisoned data, which destroys clean mapping while keeping backdoor mapping on a small set of flipped clean data. Then, the adversarial knowledge distillation is designed to reinforce clean mapping and suppress backdoor mapping through a cycle iteration mechanism between trust and punish distillations using clean and identified poisoned data. We conduct experiments to mitigate mainstream attacks on three datasets, and experimental results demonstrate that BeDKD surpasses the state-of-the-art defenses and reduces the ASR by 98% without significantly reducing the CACC.