arXiv subjects

Nenghai Yu

Publications and source records attributed to Nenghai Yu.

At least 73 records · Page 4Linked to original sources

arXiv2025

SafeGuider: Robust and Practical Content Safety Control for Text-to-Image Models

Text-to-image models have shown remarkable capabilities in generating high-quality images from natural language descriptions. However, these models are highly vulnerable to adversarial prompts, which can bypass safety measures and produce harmful content. Despite various defensive strategies, achieving robustness against attacks while maintaining practical utility in real-world applications remains a significant challenge. To address this issue, we first conduct an empirical study of the text encoder in the Stable Diffusion (SD) model, which is a widely used and representative text-to-image model. Our findings reveal that the [EOS] token acts as a semantic aggregator, exhibiting distinct distributional patterns between benign and adversarial prompts in its embedding space. Building on this insight, we introduce SafeGuider, a two-step framework designed for robust safety control without compromising generation quality. SafeGuider combines an embedding-level recognition model with a safety-aware feature erasure beam search algorithm. This integration enables the framework to maintain high-quality image generation for benign prompts while ensuring robust defense against both in-domain and out-of-domain attacks. SafeGuider demonstrates exceptional effectiveness in minimizing attack success rates, achieving a maximum rate of only 5.48\% across various attack scenarios. Moreover, instead of refusing to generate or producing black images for unsafe prompts, SafeGuider generates safe and meaningful images, enhancing its practical utility. In addition, SafeGuider is not limited to the SD model and can be effectively applied to other text-to-image models, such as the Flux model, demonstrating its versatility and adaptability across different architectures. We hope that SafeGuider can shed some light on the practical deployment of secure text-to-image systems.

cs.CR↗

arXiv2025

Flora: Effortless Context Construction to Arbitrary Length and Scale

Effectively handling long contexts is challenging for Large Language Models (LLMs) due to the rarity of long texts, high computational demands, and substantial forgetting of short-context abilities. Recent approaches have attempted to construct long contexts for instruction tuning, but these methods often require LLMs or human interventions, which are both costly and limited in length and diversity. Also, the drop in short-context performances of present long-context LLMs remains significant. In this paper, we introduce Flora, an effortless (human/LLM-free) long-context construction strategy. Flora can markedly enhance the long-context performance of LLMs by arbitrarily assembling short instructions based on categories and instructing LLMs to generate responses based on long-context meta-instructions. This enables Flora to produce contexts of arbitrary length and scale with rich diversity, while only slightly compromising short-context performance. Experiments on Llama3-8B-Instruct and QwQ-32B show that LLMs enhanced by Flora excel in three long-context benchmarks while maintaining strong performances in short-context tasks. Our data-construction code is available at \href{https://github.com/txchen-USTC/Flora}{https://github.com/txchen-USTC/Flora}.

cs.CL↗

arXiv2025

WavInWav: Time-domain Speech Hiding via Invertible Neural Network

Data hiding is essential for secure communication across digital media, and recent advances in Deep Neural Networks (DNNs) provide enhanced methods for embedding secret information effectively. However, previous audio hiding methods often result in unsatisfactory quality when recovering secret audio, due to their inherent limitations in the modeling of time-frequency relationships. In this paper, we explore these limitations and introduce a new DNN-based approach. We use a flow-based invertible neural network to establish a direct link between stego audio, cover audio, and secret audio, enhancing the reversibility of embedding and extracting messages. To address common issues from time-frequency transformations that degrade secret audio quality during recovery, we implement a time-frequency loss on the time-domain signal. This approach not only retains the benefits of time-frequency constraints but also enhances the reversibility of message recovery, which is vital for practical applications. We also add an encryption technique to protect the hidden data from unauthorized access. Experimental results on the VCTK and LibriSpeech datasets demonstrate that our method outperforms previous approaches in terms of subjective and objective metrics and exhibits robustness to various types of noise, suggesting its utility in targeted secure communication scenarios.

cs.SD↗

arXiv2025

PSRT: Accelerating LRM-based Guard Models via Prefilled Safe Reasoning Traces

Large Reasoning Models (LRMs) have demonstrated remarkable performance on tasks such as mathematics and code generation. Motivated by these strengths, recent work has empirically demonstrated the effectiveness of LRMs as guard models in improving harmful query detection. However, LRMs typically generate long reasoning traces during inference, causing substantial computational overhead. In this paper, we introduce PSRT, a method that replaces the model's reasoning process with a Prefilled Safe Reasoning Trace, thereby significantly reducing the inference cost of LRMs. Concretely, PSRT prefills "safe reasoning virtual tokens" from a constructed dataset and learns over their continuous embeddings. With the aid of indicator tokens, PSRT enables harmful-query detection in a single forward pass while preserving the classification effectiveness of LRMs. We evaluate PSRT on 7 models, 13 datasets, and 8 jailbreak methods. In terms of efficiency, PSRT completely removes the overhead of generating reasoning tokens during inference. In terms of classification performance, PSRT achieves nearly identical accuracy, with only a minor average F1 drop of 0.015 across 7 models and 5 datasets.

cs.CR↗

arXiv2025

Multimodal Prompt Decoupling Attack on the Safety Filters in Text-to-Image Models

Text-to-image (T2I) models have been widely applied in generating high-fidelity images across various domains. However, these models may also be abused to produce Not-Safe-for-Work (NSFW) content via jailbreak attacks. Existing jailbreak methods primarily manipulate the textual prompt, leaving potential vulnerabilities in image-based inputs largely unexplored. Moreover, text-based methods face challenges in bypassing the model's safety filters. In response to these limitations, we propose the Multimodal Prompt Decoupling Attack (MPDA), which utilizes image modality to separate the harmful semantic components of the original unsafe prompt. MPDA follows three core steps: firstly, a large language model (LLM) decouples unsafe prompts into pseudo-safe prompts and harmful prompts. The former are seemingly harmless sub-prompts that can bypass filters, while the latter are sub-prompts with unsafe semantics that trigger filters. Subsequently, the LLM rewrites the harmful prompts into natural adversarial prompts to bypass safety filters, which guide the T2I model to modify the base image into an NSFW output. Finally, to ensure semantic consistency between the generated NSFW images and the original unsafe prompts, the visual language model generates image captions, providing a new pathway to guide the LLM in iterative rewriting and refining the generated content.

cs.CV↗

arXiv2025

MARS-Bench: A Multi-turn Athletic Real-world Scenario Benchmark for Dialogue Evaluation

Large Language Models (\textbf{LLMs}), e.g. ChatGPT, have been widely adopted in real-world dialogue applications. However, LLMs' robustness, especially in handling long complex dialogue sessions, including frequent motivation transfer, sophisticated cross-turn dependency, is criticized all along. Nevertheless, no existing benchmarks can fully reflect these weaknesses. We present \textbf{MARS-Bench}, a \textbf{M}ulti-turn \textbf{A}thletic \textbf{R}eal-world \textbf{S}cenario Dialogue \textbf{Bench}mark, designed to remedy the gap. MARS-Bench is constructed from play-by-play text commentary so to feature realistic dialogues specifically designed to evaluate three critical aspects of multi-turn conversations: Ultra Multi-turn, Interactive Multi-turn, and Cross-turn Tasks. Extensive experiments on MARS-Bench also reveal that closed-source LLMs significantly outperform open-source alternatives, explicit reasoning significantly boosts LLMs' robustness on handling long complex dialogue sessions, and LLMs indeed face significant challenges when handling motivation transfer and sophisticated cross-turn dependency. Moreover, we provide mechanistic interpretability on how attention sinks due to special tokens lead to LLMs' performance degradation when handling long complex dialogue sessions based on attention visualization experiment in Qwen2.5-7B-Instruction.

cs.CL↗

arXiv2025

Scale Your Instructions: Enhance the Instruction-Following Fidelity of Unified Image Generation Model by Self-Adaptive Attention Scaling

Recent advancements in unified image generation models, such as OmniGen, have enabled the handling of diverse image generation and editing tasks within a single framework, accepting multimodal, interleaved texts and images in free form. This unified architecture eliminates the need for text encoders, greatly reducing model complexity and standardizing various image generation and editing tasks, making it more user-friendly. However, we found that it suffers from text instruction neglect, especially when the text instruction contains multiple sub-instructions. To explore this issue, we performed a perturbation analysis on the input to identify critical steps and layers. By examining the cross-attention maps of these key steps, we observed significant conflicts between neglected sub-instructions and the activations of the input image. In response, we propose Self-Adaptive Attention Scaling (SaaS), a method that leverages the consistency of cross-attention between adjacent timesteps to dynamically scale the attention activation for each sub-instruction. Our SaaS enhances instruction-following fidelity without requiring additional training or test-time optimization. Experimental results on instruction-based image editing and visual conditional image generation validate the effectiveness of our SaaS, showing superior instruction-following fidelity over existing methods. The code is available https://github.com/zhouchao-ops/SaaS.

cs.CV↗

arXiv2025

BURN: Backdoor Unlearning via Adversarial Boundary Analysis

Backdoor unlearning aims to remove backdoor-related information while preserving the model's original functionality. However, existing unlearning methods mainly focus on recovering trigger patterns but fail to restore the correct semantic labels of poison samples. This limitation prevents them from fully eliminating the false correlation between the trigger pattern and the target label. To address this, we leverage boundary adversarial attack techniques, revealing two key observations. First, poison samples exhibit significantly greater distances from decision boundaries compared to clean samples, indicating they require larger adversarial perturbations to change their predictions. Second, while adversarial predicted labels for clean samples are uniformly distributed, those for poison samples tend to revert to their original correct labels. Moreover, the features of poison samples restore to closely resemble those of corresponding clean samples after adding adversarial perturbations. Building upon these insights, we propose Backdoor Unlearning via adversaRial bouNdary analysis (BURN), a novel defense framework that integrates false correlation decoupling, progressive data refinement, and model purification. In the first phase, BURN employs adversarial boundary analysis to detect poisoned samples based on their abnormal adversarial boundary distances, then restores their correct semantic labels for fine-tuning. In the second phase, it employs a feedback mechanism that tracks prediction discrepancies between the original backdoored model and progressively sanitized models, guiding both dataset refinement and model purification. Extensive evaluations across multiple datasets, architectures, and seven diverse backdoor attack types confirm that BURN effectively removes backdoor threats while maintaining the model's original performance.

cs.CR↗

arXiv2025

De-AntiFake: Rethinking the Protective Perturbations Against Voice Cloning Attacks

The rapid advancement of speech generation models has heightened privacy and security concerns related to voice cloning (VC). Recent studies have investigated disrupting unauthorized voice cloning by introducing adversarial perturbations. However, determined attackers can mitigate these protective perturbations and successfully execute VC. In this study, we conduct the first systematic evaluation of these protective perturbations against VC under realistic threat models that include perturbation purification. Our findings reveal that while existing purification methods can neutralize a considerable portion of the protective perturbations, they still lead to distortions in the feature space of VC models, which degrades the performance of VC. From this perspective, we propose a novel two-stage purification method: (1) Purify the perturbed speech; (2) Refine it using phoneme guidance to align it with the clean speech distribution. Experimental results demonstrate that our method outperforms state-of-the-art purification methods in disrupting VC defenses. Our study reveals the limitations of adversarial perturbation-based VC defenses and underscores the urgent need for more robust solutions to mitigate the security and privacy risks posed by VC. The code and audio samples are available at https://de-antifake.github.io.

cs.SD↗

arXiv2025

ScaleCap: Inference-Time Scalable Image Captioning via Dual-Modality Debiasing

This paper presents ScaleCap, an inference-time scalable image captioning strategy that generates comprehensive and detailed image captions. The key challenges of high-quality image captioning lie in the inherent biases of LVLMs: multimodal bias resulting in imbalanced descriptive granularity, offering detailed accounts of some elements while merely skimming over others; linguistic bias leading to hallucinated descriptions of non-existent objects. To address these issues, we propose a scalable debiased captioning strategy, which continuously enriches and calibrates the caption with increased inference budget. Specifically, we propose two novel components: heuristic question answering and contrastive sentence rating. The former generates content-specific questions based on the image and answers them to progressively inject relevant information into the caption. The latter employs sentence-level offline contrastive decoding to effectively identify and eliminate hallucinations caused by linguistic biases. With increased inference cost, more heuristic questions are raised by ScaleCap to progressively capture additional visual details, generating captions that are more accurate, balanced, and informative. Extensive modality alignment experiments demonstrate the effectiveness of ScaleCap. Annotating 450K images with ScaleCap and using them for LVLM pretraining leads to consistent performance gains across 11 widely used benchmarks. Furthermore, ScaleCap showcases superb richness and fidelity of generated captions with two additional tasks: replacing images with captions in VQA task, and reconstructing images from captions to assess semantic coverage. Code is available at https://github.com/Cooperx521/ScaleCap.

cs.CV↗

arXiv2025

M3-AGIQA: Multimodal, Multi-Round, Multi-Aspect AI-Generated Image Quality Assessment

The rapid advancement of AI-generated image (AIGI) models presents new challenges for evaluating image quality, particularly across three aspects: perceptual quality, prompt correspondence, and authenticity. To address these challenges, we introduce M3-AGIQA, a comprehensive framework that leverages Multimodal Large Language Models (MLLMs) to enable more human-aligned, holistic evaluation of AI-generated images across both visual and textual domains. Besides, our framework features a structured multi-round evaluation process, generating and analyzing intermediate image descriptions to provide deeper insight into these three aspects. By aligning model outputs more closely with human judgment, M3-AGIQA delivers robust and interpretable quality scores. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art performance on tested datasets and aspects, and exhibits strong generalizability in most cross-dataset settings. Code is available at https://github.com/strawhatboy/M3-AGIQA.

cs.CV↗

arXiv2025

SQL Injection Jailbreak: A Structural Disaster of Large Language Models

Large Language Models (LLMs) are susceptible to jailbreak attacks that can induce them to generate harmful content. Previous jailbreak methods primarily exploited the internal properties or capabilities of LLMs, such as optimization-based jailbreak methods and methods that leveraged the model's context-learning abilities. In this paper, we introduce a novel jailbreak method, SQL Injection Jailbreak (SIJ), which targets the external properties of LLMs, specifically, the way LLMs construct input prompts. By injecting jailbreak information into user prompts, SIJ successfully induces the model to output harmful content. For open-source models, SIJ achieves near 100% attack success rates on five well-known LLMs on the AdvBench and HEx-PHI, while incurring lower time costs compared to previous methods. For closed-source models, SIJ achieves an average attack success rate over 85% across five models in the GPT and Doubao series. Additionally, SIJ exposes a new vulnerability in LLMs that urgently requires mitigation. To address this, we propose a simple adaptive defense method called Self-Reminder-Key to counter SIJ and demonstrate its effectiveness through experimental results. Our code is available at https://github.com/weiyezhimeng/SQL-Injection-Jailbreak.

cs.CR↗

arXiv2025

Gaussian Shading++: Rethinking the Realistic Deployment Challenge of Performance-Lossless Image Watermark for Diffusion Models

Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. Existing methods primarily focus on ensuring that watermark embedding does not degrade the model performance. However, they often overlook critical challenges in real-world deployment scenarios, such as the complexity of watermark key management, user-defined generation parameters, and the difficulty of verification by arbitrary third parties. To address this issue, we propose Gaussian Shading++, a diffusion model watermarking method tailored for real-world deployment. We propose a double-channel design that leverages pseudorandom error-correcting codes to encode the random seed required for watermark pseudorandomization, achieving performance-lossless watermarking under a fixed watermark key and overcoming key management challenges. Additionally, we model the distortions introduced during generation and inversion as an additive white Gaussian noise channel and employ a novel soft decision decoding strategy during extraction, ensuring strong robustness even when generation parameters vary. To enable third-party verification, we incorporate public key signatures, which provide a certain level of resistance against forgery attacks even when model inversion capabilities are fully disclosed. Extensive experiments demonstrate that Gaussian Shading++ not only maintains performance losslessness but also outperforms existing methods in terms of robustness, making it a more practical solution for real-world deployment.

cs.CV↗

arXiv2025

BinMetric: A Comprehensive Binary Analysis Benchmark for Large Language Models

Binary analysis remains pivotal in software security, offering insights into compiled programs without source code access. As large language models (LLMs) continue to excel in diverse language understanding and generation tasks, their potential in decoding complex binary data structures becomes evident. However, the lack of standardized benchmarks in this domain limits the assessment and comparison of LLM's capabilities in binary analysis and hinders the progress of research and practical applications. To bridge this gap, we introduce BinMetric, a comprehensive benchmark designed specifically to evaluate the performance of large language models on binary analysis tasks. BinMetric comprises 1,000 questions derived from 20 real-world open-source projects across 6 practical binary analysis tasks, including decompilation, code summarization, assembly instruction generation, etc., which reflect actual reverse engineering scenarios. Our empirical study on this benchmark investigates the binary analysis capabilities of various state-of-the-art LLMs, revealing their strengths and limitations in this field. The findings indicate that while LLMs show strong potential, challenges still exist, particularly in the areas of precise binary lifting and assembly synthesis. In summary, BinMetric makes a significant step forward in measuring the binary analysis capabilities of LLMs, establishing a new benchmark leaderboard, and our study provides valuable insights for the future development of these LLMs in software security.

cs.SE↗

arXiv2025

CompileAgent: Automated Real-World Repo-Level Compilation with Tool-Integrated LLM-based Agent System

With open-source projects growing in size and complexity, manual compilation becomes tedious and error-prone, highlighting the need for automation to improve efficiency and accuracy. However, the complexity of compilation instruction search and error resolution makes automatic compilation challenging. Inspired by the success of LLM-based agents in various fields, we propose CompileAgent, the first LLM-based agent framework dedicated to repo-level compilation. CompileAgent integrates five tools and a flow-based agent strategy, enabling interaction with software artifacts for compilation instruction search and error resolution. To measure the effectiveness of our method, we design a public repo-level benchmark CompileAgentBench, and we also design two baselines for comparison by combining two compilation-friendly schemes. The performance on this benchmark shows that our method significantly improves the compilation success rate, ranging from 10% to 71%. Meanwhile, we evaluate the performance of CompileAgent under different agent strategies and verify the effectiveness of the flow-based strategy. Additionally, we emphasize the scalability of CompileAgent, further expanding its application prospects.

cs.SE↗

arXiv2025

On the Vulnerability of Text Sanitization

Text sanitization, which employs differential privacy to replace sensitive tokens with new ones, represents a significant technique for privacy protection. Typically, its performance in preserving privacy is evaluated by measuring the attack success rate (ASR) of reconstruction attacks, where attackers attempt to recover the original tokens from the sanitized ones. However, current reconstruction attacks on text sanitization are developed empirically, making it challenging to accurately assess the effectiveness of sanitization. In this paper, we aim to provide a more accurate evaluation of sanitization effectiveness. Inspired by the works of Palamidessi et al., we implement theoretically optimal reconstruction attacks targeting text sanitization. We derive their bounds on ASR as benchmarks for evaluating sanitization performance. For real-world applications, we propose two practical reconstruction attacks based on these theoretical findings. Our experimental results underscore the necessity of reassessing these overlooked risks. Notably, one of our attacks achieves a 46.4% improvement in ASR over the state-of-the-art baseline, with a privacy budget of epsilon=4.0 on the SST-2 dataset. Our code is available at: https://github.com/mengtong0110/On-the-Vulnerability-of-Text-Sanitization.

cs.CR↗

arXiv2025

An Empirical Study on the Effectiveness of Large Language Models for Binary Code Understanding

Binary code analysis plays a pivotal role in the field of software security and is widely used in tasks such as software maintenance, malware detection, software vulnerability discovery, patch analysis, etc. However, unlike source code, reverse engineers face significant challenges in understanding binary code due to the lack of intuitive semantic information. Although traditional reverse tools can convert binary code into C-like pseudo code, the lack of code comments and symbolic information such as function names still makes code understanding difficult. In recent years, two groups of techniques have shown promising prospects: (1) Deep learning-based techniques have demonstrated competitive results in tasks related to binary code understanding, furthermore, (2) Large Language Models (LLMs) have been extensively pre-trained at the source-code level for tasks such as code understanding and generation. This has left participants wondering about the capabilities of LLMs in binary code understanding. To this end, this work proposes a benchmark to evaluate the effectiveness of LLMs in real-world reverse engineering scenarios, which covers two key binary code understanding tasks, i.e., function name recovery and binary code summarization. To more comprehensively evaluate, we include binaries with multiple target architectures as well as different optimization options. We gain valuable insights into the capabilities and limitations through extensive empirical studies of popular LLMs using our benchmark. Our evaluations reveal that existing LLMs can understand binary code to a certain extent, thereby improving the efficiency of binary code analysis. Our results highlight the great potential of the LLMs in advancing the field of binary code understanding, and provide new directions for binary code analysis techniques.

cs.SE↗

arXiv2025

Provably Secure Public-Key Steganography Based on Admissible Encoding

The technique of hiding secret messages within seemingly harmless covertext to evade examination by censors with rigorous security proofs is known as provably secure steganography (PSS). PSS evolves from symmetric key steganography to public-key steganography, functioning without the requirement of a pre-shared key and enabling the extension to multi-party covert communication and identity verification mechanisms. Recently, a public-key steganography method based on elliptic curves was proposed, which uses point compression to eliminate the algebraic structure of curve points. However, this method has strict requirements on the curve parameters and is only available on half of the points. To overcome these limitations, this paper proposes a more general elliptic curve public key steganography method based on admissible encoding. By applying the tensor square function to the known well-distributed encoding, we construct admissible encoding, which can create the pseudo-random public-key encryption function. The theoretical analysis and experimental results show that the proposed provable secure public-key steganography method can be deployed on all types of curves and utilize all points on the curve.

cs.CR↗