论文检索

Minseok Choi, Daniel Rim, Dohyun Lee, Jaegul Choo

Instruction-following large language models (LLMs), such as ChatGPT, have become widely popular among everyday users. However, these models inadvertently disclose private, sensitive information to their users, underscoring the need for machine unlearning techniques to remove selective information from the models. While prior work has focused on forgetting small, random subsets of training data at the instance-level, we argue that real-world scenarios often require the removal of an entire user data, which may require a more careful maneuver. In this study, we explore entity-level unlearning, which aims to erase all knowledge related to a target entity while preserving the remaining model capabilities. To address this, we introduce Opt-Out, an optimal transport-based unlearning method that utilizes the Wasserstein distance from the model’s initial parameters to achieve more effective and fine-grained unlearning. We also present the first Entity-Level Unlearning Dataset (ELUDe) designed to evaluate entity-level unlearning. Our empirical results demonstrate that Opt-Out surpasses existing methods, establishing a new standard for secure and adaptable LLMs that can accommodate user data removal requests without the need for full retraining.

Haoyang Li, Huan Gao, Zhiyuan Zhao, Zhiyu Lin, Junyu Gao, Xuelong Li

The widespread adoption of Large Language Models (LLMs) has heightened concerns about their security, particularly their vulnerability to jailbreak attacks that leverage crafted prompts to generate malicious outputs. While prior research has been conducted on general security capabilities of LLMs, their specific susceptibility to jailbreak attacks in code generation remains largely unexplored. To fill this gap, we propose MalwareBench, a benchmark dataset containing 3,520 jailbreaking prompts for malicious code-generation, designed to evaluate LLM robustness against such threats. MalwareBench is based on 320 manually crafted malicious code generation requirements, covering 11 jailbreak methods and 29 code functionality categories. Experiments show that mainstream LLMs exhibit limited ability to reject malicious code-generation requirements, and the combination of multiple jailbreak methods further reduces the model’s security capabilities: specifically, the average rejection rate for malicious content is 60.93%, dropping to 39.92% when combined with jailbreak attack algorithms. Our work highlights that the code security capabilities of LLMs still pose significant challenges.

Weiyang Guo, Jing Li, Wenya Wang, Yu Li, Daojing He, Jun Yu, Min Zhang

The proliferation of jailbreak attacks against large language models (LLMs) highlights the need for robust security measures. However, in multi-round dialogues, malicious intentions may be hidden in interactions, leading LLMs to be more prone to produce harmful responses. In this paper, we propose the Multi-Turn Safety Alignment (MTSA) framework, to address the challenge of securing LLMs in multi-round interactions. It consists of two stages: In the thought-guided attack learning stage, the red-team model learns about thought-guided multi-round jailbreak attacks to generate adversarial prompts. In the adversarial iterative optimization stage, the red-team model and the target model continuously improve their respective capabilities in interaction. Furthermore, we introduce a multi-turn reinforcement learning algorithm based on future rewards to enhance the robustness of safety alignment. Experimental results show that the red-team model exhibits state-of-the-art attack capabilities, while the target model significantly improves its performance on safety benchmarks.

Yuxuan Zhou, Margret Keuper, Mario Fritz

Sampling-based decoding strategies have been widely adopted for Large Language Models (LLMs) in numerous applications, targeting a balance between diversity and quality via temperature tuning and tail truncation. Considering the strong dependency of the candidate next tokens on different prefixes, recent studies propose to adaptively truncate the tail of LLMs’ predicted distribution. Although improved results have been reported with these methods on open-ended text generation tasks, the results are highly dependent on the curated parameters and the limited exemplar text. In this paper, we propose a systematic way to estimate the intrinsic capacity of a truncation sampling method by considering the trade-off between diversity and risk at each decoding step, based on our collected prefix tree which preserves the context of a full sentence. Our work offers a comprehensive comparison of existing truncation sampling methods and serves as a practical user guideline for their parameter selection. Our code is available at https://anonymous.4open.science/r/Truncation-Sampling-Evaluation-251F.

Mohammad Saqib Hasan, Saikat Chakraborty, Santu Karmaker, Niranjan Balasubramanian

LLM generated code often contains security issues. We address two key challenges in improving secure code generation. First, obtaining high quality training data covering a broad set of security issues is critical. To address this, we introduce a method for distilling a preference dataset of insecure and secure code pairs from frontier LLMs, along with a security reasoning that explains the issues and the fix. The key idea here is to make use of security knowledge sources to devise a systematic prompting strategy that ensures broad coverage. Second, aligning models to secure code requires focusing on localized regions of code. Direct preference optimization methods, like SimPO, are not designed to handle these localized differences and turn out to be ineffective. We address this with a new localized preference optimization algorithm that masks the security related tokens in both the winning (secure) and losing (insecure) responses. To prevent loss in code quality, we also add a regularizer. Evaluations show that both training on our dataset, DiSCo, and the new preference optimization algorithm, LPO, yield substantial reductions in code insecurity while also improving overall code quality. Code and dataset are available at https://github.com/StonyBrookNLP/disco-lpo.

Lang Gao, Jiahui Geng, Xiangliang Zhang, Preslav Nakov, Xiuying Chen

Jailbreaking in Large Language Models (LLMs) is a major security concern as it can deceive LLMs into generating harmful text. However, understanding of how jailbreaking works remains limited, hindering the development of effective defense strategies. To address this issue, we conduct a large-scale analysis of seven different jailbreak methods and identify that disagreements among methods stem from insufficient observation samples.We introduce the concept of a safety boundary and discover that jailbreaks shift harmful activations outside this boundary, where LLMs become less sensitive to harmful information. Our analysis reveals that low and middle layers play a critical role in these shifts, while deeper layers have a lesser impact.Building on these insights, we propose a novel defense mechanism called Activation Boundary Defense (ABD), which adaptively constrains activations within the safety boundary. To enhance its effectiveness, we use Bayesian optimization to selectively apply the defense to the low and middle layers.Experiments on several benchmark datasets demonstrate that ABD achieves an average Defense Success Rate (DSR) of over 98% against various jailbreak attacks, with less than a 2% impact on the model’s general capabilities.

Weikai Lu, Hao Peng, Huiping Zhuang, Cen Chen, Ziqian Zeng

Multimodal Large Language Models (MLLMs) have serious security vulnerabilities. While safety alignment using multimodal datasets consisting of text and data of additional modalities can effectively enhance MLLM’s security, it is costly to construct these datasets. Existing low-resource security alignment methods, including textual alignment, have been found to struggle with the security risks posed by additional modalities. To address this, we propose Synthetic Embedding augmented safety Alignment (SEA), which optimizes embeddings of additional modality through gradient updates to expand textual datasets. This enables multimodal safety alignment training even when only textual data is available. Extensive experiments on image, video, and audio-based MLLMs demonstrate that SEA can synthesize a high-quality embedding on a single RTX3090 GPU within 24 seconds. SEA significantly improves the security of MLLMs when faced with threats from additional modalities. To assess the security risks introduced by video and audio, we also introduced a new benchmark called VA-SafetyBench. High attack success rates across multiple MLLMs validate its challenge. Our code and data will be available at https://github.com/ZeroNLP/SEA.

Weixiang Zhao, Yulin Hu, Yang Deng, Tongtong Wu, Wenxuan Zhang, Jiahe Guo, An Zhang, Yanyan Zhao, Bing Qin, Tat-Seng Chua 等

Large language models (LLMs) have become increasingly central to AI applications worldwide, necessitating robust multilingual safety alignment to ensure secure deployment across diverse linguistic contexts. Existing preference learning methods for safety alignment, such as RLHF and DPO, are primarily monolingual and struggle with noisy multilingual data. To address these limitations, we introduce Multilingual reward gaP Optimization (MPO), a novel approach that leverages the well-aligned safety capabilities of the dominant language (e.g., English) to improve safety alignment across multiple languages. MPO directly minimizes the reward gap difference between the dominant language and target languages, effectively transferring safety capabilities while preserving the original strengths of the dominant language. Extensive experiments on three LLMs, LLaMA-3.1, Gemma-2 and Qwen2.5, validate MPO’s efficacy in multilingual safety alignment without degrading general multilingual utility.

Zhen Sun, Zongmin Zhang, Xinyue Shen, Ziyi Zhang, Yule Liu, Michael Backes, Yang Zhang, Xinlei He

Social media platforms are experiencing a growing presence of AI-Generated Texts (AIGTs). However, the misuse of AIGTs could have profound implications for public opinion, such as spreading misinformation and manipulating narratives. Despite its importance, it remains unclear how prevalent AIGTs are on social media. To address this gap, this paper aims to quantify and monitor the AIGTs on online social media platforms. We first collect a dataset (SM-D) with around 2.4M posts from 3 major social media platforms: Medium, Quora, and Reddit. Then, we construct a diverse dataset (AIGTBench) to train and evaluate AIGT detectors. AIGTBench combines popular open-source datasets and our AIGT datasets generated from social media texts by 12 LLMs, serving as a benchmark for evaluating mainstream detectors. With this setup, we identify the best-performing detector (OSM-Det). We then apply OSM-Det to SM-D to track AIGTs across social media platforms from January 2022 to October 2024, using the AI Attribution Rate (AAR) as the metric. Specifically, Medium and Quora exhibit marked increases in AAR, rising from 1.77% to 37.03% and 2.06% to 38.95%, respectively. In contrast, Reddit shows slower growth, with AAR increasing from 1.31% to 2.45% over the same period. Our further analysis indicates that AIGTs on social media differ from human-written texts across several dimensions, including linguistic patterns, topic distributions, engagement levels, and the follower distribution of authors. We envision our analysis and findings on AIGTs in social media can shed light on future research in this domain.

Jinglong Luo, Guanzhong Chen, Yehong Zhang, Shiyu Liu, Hui Wang, Yue Yu, Xun Zhou, Yuan Qi, Zenglin Xu

With the growing deployment of pre-trained models like Transformers on cloud platforms, privacy concerns about model parameters and inference data are intensifying. Existing Privacy-Preserving Transformer Inference (PPTI) frameworks face the “impossible trinity” of balancing privacy, efficiency, and performance: Secure Multi-Party Computation (SMPC)-based approaches ensure strong privacy but suffer from high computational overhead and performance losses; Conversely, permutation-based methods achieve near-plaintext efficiency and accuracy but compromise privacy by exposing sensitive model parameters and intermediate results. Bridging this gap with a single approach presents substantial challenges, motivating the introduction of CENTAUR, a groundbreaking PPTI framework that seamlessly integrates random permutations and SMPC to address the “impossible trinity”. By designing efficient PPTI algorithms tailored to the structural properties of Transformer models, CENTAUR achieves an unprecedented balance among privacy, efficiency, and performance. Our experiments demonstrate CENTAUR’s ability to resist diverse data reconstruction attacks, achieve plaintext-level inference accuracy, and boost inference speed by 5.0~30.4 times, unlocking new possibilities for secure and efficient AI deployment.

Junjie Chu, Yugeng Liu, Ziqing Yang, Xinyue Shen, Michael Backes, Yang Zhang

Jailbreak attacks aim to bypass the LLMs’ safeguards. While researchers have proposed different jailbreak attacks in depth, they have done so in isolation—either with unaligned settings or comparing a limited range of methods. To fill this gap, we present a large-scale evaluation of various jailbreak attacks. We collect 17 representative jailbreak attacks, summarize their features, and establish a novel jailbreak attack taxonomy. Then we conduct comprehensive measurement and ablation studies across nine aligned LLMs on 160 forbidden questions from 16 violation categories. Also, we test jailbreak attacks under eight advanced defenses. Based on our taxonomy and experiments, we identify some important patterns, such as heuristic-based attacks, which could achieve high attack success rates but are easy to mitigate by defenses. Our study offers valuable insights for future research on jailbreak attacks and defenses and serves as a benchmark tool for researchers and practitioners to evaluate them effectively.

Xiaoyu Zhang, Juan Zhai, Shiqing Ma, Qingshuang Bao, Weipeng Jiang, Qian Wang, Chao Shen, Yang Liu

Large Language Models (LLMs) have emerged as the new recommendation engines, surpassing traditional methods in both capability and scope, particularly in code generation. In this paper, we reveal a novel **provider bias** in LLMs: without explicit directives, these models show systematic preferences for services from specific providers in their recommendations (e.g., favoring Google Cloud over Microsoft Azure). To systematically investigate this bias, we develop an automated pipeline to construct the dataset, incorporating 6 distinct coding task categories and 30 real-world application scenarios. Leveraging this dataset, we conduct the **first** comprehensive empirical study of provider bias in LLM code generation across seven state-of-the-art LLMs, utilizing approximately 500 million tokens (equivalent to $5,000+ in computational costs). Our findings reveal that LLMs exhibit significant provider preferences, predominantly favoring services from Google and Amazon, and can autonomously modify input code to incorporate their preferred providers without users’ requests. Such a bias holds far-reaching implications for market dynamics and societal equilibrium, potentially contributing to digital monopolies. It may also deceive users and violate their expectations, leading to various consequences. We call on the academic community to recognize this emerging issue and develop effective evaluation and mitigation methods to uphold AI security and fairness.

Xinyue Shen, Yun Shen, Michael Backes, Yang Zhang

Knowledge files have been widely used in large language model (LLM)-powered agents, such as GPTs, to improve response quality. However, concerns over the potential leakage of knowledge files have grown significantly. Existing studies demonstrate that adversarial prompts can induce GPTs to leak knowledge file content. Yet, it remains uncertain whether additional leakage vectors exist, particularly given the complex data flows across clients, servers, and databases in GPTs. In this paper, we present a comprehensive risk assessment of knowledge file leakage, leveraging a novel workflow inspired by Data Security Posture Management (DSPM). Through the analysis of 651,022 GPT metadata, 11,820 flows, and 1,466 responses, we identify five leakage vectors: metadata, GPT initialization, retrieval, sandboxed execution environments, and prompts. These vectors enable adversaries to extract sensitive knowledge file data such as titles, content, types, and sizes. Notably, the activation of the built-in tool Code Interpreter leads to a privilege escalation vulnerability, enabling adversaries to directly download original knowledge files with a 95.95% success rate. Further analysis reveals that 28.80% of leaked files are copyrighted, including digital copies from major publishers and internal materials from a listed company. In the end, we provide actionable solutions for GPT builders and platform providers to secure the GPT data supply chain.

Yutao Mou, Xiao Deng, Yuxiao Luo, Shikun Zhang, Wei Ye

Code security and usability are both essential for various coding assistant applications driven by large language models (LLMs). Current code security benchmarks focus solely on single evaluation task and paradigm, such as code completion and generation, lacking comprehensive assessment across dimensions like secure code generation, vulnerability repair and discrimination. In this paper, we first propose CoV-Eval, a multi-task benchmark covering various tasks such as code completion, vulnerability repair, vulnerability detection and classification, for comprehensive evaluation of LLM code security. Besides, we developed VC-Judge, an improved judgment model that aligns closely with human experts and can review LLM-generated programs for vulnerabilities in a more efficient and reliable way. We conduct a comprehensive evaluation of 20 proprietary and open-source LLMs. Overall, while most LLMs identify vulnerable codes well, they still tend to generate insecure codes and struggle with recognizing specific vulnerability types and performing repairs. Extensive experiments and qualitative analyses reveal key challenges and optimization directions, offering insights for future research in LLM code security.

Jongwook Han, Dongmin Choi, Woojung Song, Eun-Ju Lee, Yohan Jo

The importance of benchmarks for assessing the values of language models has been pronounced due to the growing need of more authentic, human-aligned responses. However, existing benchmarks rely on human or machine annotations that are vulnerable to value-related biases. Furthermore, the tested scenarios often diverge from real-world contexts in which models are commonly used to generate text and express values. To address these issues, we propose the Value Portrait benchmark, a reliable framework for evaluating LLMs’ value orientations with two key characteristics. First, the benchmark consists of items that capture real-life user-LLM interactions, enhancing the relevance of assessment results to real-world LLM usage. Second, each item is rated by human subjects based on its similarity to their own thoughts, and correlations between these ratings and the subjects’ actual value scores are derived. This psychometrically validated approach ensures that items strongly correlated with specific values serve as reliable items for assessing those values. Through evaluating 44 LLMs with our benchmark, we find that these models prioritize Benevolence, Security, and Self-Direction values while placing less emphasis on Tradition, Power, and Achievement values. Also, our analysis reveals biases in how LLMs perceive various demographic groups, deviating from real human data.

Zihan Liu, Yizhen Wang, Rui Wang, Sai Wu

Integrating split learning with large language model fine-tuning (LLM-FT) enables secure collaboration between a trusted local client and a well-equipped remote server, but it is vulnerable to data reconstruction attacks (DRAs) that exploit transmitted activations and gradients. Current defense methods, like adding noise to activations or gradients, often sacrifice task-specific model performance under strict privacy constraints. This paper introduces DualGuard, a bidirectional defense mechanism against DRAs for split-based LLM-FT. DualGuard proposes a local warm-up parameter space transformation to alter client-side model parameters before training, using multi-task learning to strike a balance between privacy protection and model performance. Additionally, a global fine-tuning parameter space retention strategy prevents the model from reverting to vulnerable states during formal fine-tuning. Experiments show that DualGuard outperforms current defense methods against various DRAs, while maintaining task performance. Our code will be made publicly available.

Chenxi Dai, Lin Lu, Pan Zhou

Decentralized training has become a resource-efficient framework to democratize the training of large language models (LLMs). However, the privacy risks associated with this framework, particularly due to the potential inclusion of sensitive data in training datasets, remain unexplored. This paper identifies a novel and realistic attack surface: the privacy leakage from training data in decentralized training, and proposes \textit{activation inversion attack} (AIA) for the first time. AIA first constructs a shadow dataset comprising text labels and corresponding activations using public datasets. Leveraging this dataset, an attack model can be trained to reconstruct the training data from activations in victim decentralized training. We conduct extensive experiments on various LLMs and publicly available datasets to demonstrate the susceptibility of decentralized training to AIA. These findings highlight the urgent need to enhance security measures in decentralized training to mitigate privacy risks in training LLMs.

Xindi Li, Zhe Liu, Tong Zhang, Jiahao Chen, Qingming Li, Jinbao Li, Shouling Ji

Text-to-image (T2I) models excel at generating high-quality images from text via powerful text encoders but training these encoders demands substantial computational resources. Consequently, many users seek pre-trained text encoders from model plugin-sharing platforms like Civitai and Hugging Face, which introduces an underexplored threat: the potential for adversaries to embed Trojans within these plugins. Existing Trojan attacks often require extensive training data and suffer from poor generalization across different triggers, limiting their effectiveness and scalability. To the best of our knowledge, this paper introduces the first **T**ext-encoder **W**eight-editing method for **I**nserting **S**ecret **T**rojans (**TWIST**). By identifying the *bottleneck MLP layer*—the critical point where minimal edits can dominantly control cross-modal alignment—TWIST achieves training-free and data-free Trojan insertion, which makes it highly efficient and practical. The experimental results across various triggers demonstrate that TWIST attains an average attack success rate of 91%, a 78% improvement over the state-of-the-art (SOTA) method proposed in 2024 and highlights the excellent generalization capability. Moreover, TWIST reduces modified parameters by 8-fold and cuts injection time to 25 seconds. Our findings underscore the security risks associated with text encoders in real-world applications and emphasize the need for more robust defense mechanisms.

Yidan Wang, Yubing Ren, Yanan Cao, Binxing Fang

The rise of Large Language Models (LLMs) has heightened concerns about the misuse of AI-generated text, making watermarking a promising solution. Mainstream watermarking schemes for LLMs fall into two categories: logits-based and sampling-based. However, current schemes entail trade-offs among robustness, text quality, and security. To mitigate this, we integrate logits-based and sampling-based schemes, harnessing their respective strengths to achieve synergy. In this paper, we propose a versatile symbiotic watermarking framework with three strategies: serial, parallel, and hybrid. The hybrid framework adaptively embeds watermarks using token entropy and semantic entropy, optimizing the balance between detectability, robustness, text quality, and security. Furthermore, we validate our approach through comprehensive experiments on various datasets and models. Experimental results indicate that our method outperforms existing baselines and achieves state-of-the-art (SOTA) performance. We believe this framework provides novel insights into diverse watermarking paradigms.

Weidi Luo, Shenghong Dai, Xiaogeng Liu, Suman Banerjee, Huan Sun, Muhao Chen, Chaowei Xiao

The rapid advancements in Large Language Models (LLMs) have enabled their deployment as autonomous agents for handling complex tasks in dynamic environments. These LLMs demonstrate strong problem-solving capabilities and adaptability to multifaceted scenarios. However, their use as agents also introduces significant risks, including task-specific risks, which are identified by the agent administrator based on the specific task requirements and constraints, and systemic risks, which stem from vulnerabilities in their design or interactions, potentially compromising confidentiality, integrity, or availability (CIA) of information and triggering security risks. Existing defense agencies fail to adaptively and effectively mitigate these risks. In this paper, we propose AGrail, a lifelong agent guardrail to enhance LLM agent safety, which features adaptive safety check generation, effective safety check optimization, and tool compatibility & flexibility. Extensive experiments demonstrate that AGrail not only achieves strong performance against task-specific and system risks but also exhibits transferability across different LLM agents’ tasks.