Shengfang Zhai

Xiaoyi Chen, Yinpeng Dong, Zeyu Sun et al.

Although Deep Neural Network (DNN) has led to unprecedented progress in various natural language processing (NLP) tasks, research shows that deep models are extremely vulnerable to backdoor attacks. The existing backdoor attacks mainly inject a small number of poisoned samples into the training dataset with the labels changed to the target one. Such mislabeled samples would raise suspicion upon human inspection, potentially revealing the attack. To improve the stealthiness of textual backdoor attacks, we propose the first clean-label framework Kallima for synthesizing mimesis-style backdoor samples to develop insidious textual backdoor attacks. We modify inputs belonging to the target class with adversarial perturbations, making the model rely more on the backdoor trigger. Our framework is compatible with most existing backdoor triggers. The experimental results on three benchmark datasets demonstrate the effectiveness of the proposed method.

Yanpei Guo, Wenjie Qu, Linyu Wu et al.

Commercial large language models are typically deployed as black-box API services, requiring users to trust providers to execute inference correctly and report token usage honestly. We present IMMACULATE, a practical auditing framework that detects economically motivated deviations-such as model substitution, quantization abuse, and token overbilling-without trusted hardware or access to model internals. IMMACULATE selectively audits a small fraction of requests using verifiable computation, achieving strong detection guarantees while amortizing cryptographic overhead. Experiments on dense and MoE models show that IMMACULATE reliably distinguishes benign and malicious executions with under 1% throughput overhead. Our code is published at https://github.com/guo-yanpei/Immaculate.

Shengfang Zhai, Qingni Shen, Xiaoyi Chen et al.

At present, backdoor attacks attract attention as they do great harm to deep learning models. The adversary poisons the training data making the model being injected with a backdoor after being trained unconsciously by victims using the poisoned dataset. In the field of text, however, existing works do not provide sufficient defense against backdoor attacks. In this paper, we propose a Noise-augmented Contrastive Learning (NCL) framework to defend against textual backdoor attacks when training models with untrustworthy data. With the aim of mitigating the mapping between triggers and the target label, we add appropriate noise perturbing possible backdoor triggers, augment the training dataset, and then pull homology samples in the feature space utilizing contrastive learning objective. Experiments demonstrate the effectiveness of our method in defending three types of textual backdoor attacks, outperforming the prior works.

Shengfang Zhai, Xiaoyang Ji, Yuling Shi et al.

Diffusion language models (DLMs) have recently emerged as an alternative modeling paradigm to autoregressive (AR) language models, enabling parallel generation and bidirectional context modeling. Yet their security implications, particularly their vulnerability to backdoor attacks, remain underexplored. We propose BadDLM, a unified framework for studying backdoor attacks against DLMs with diverse targets. We introduce a trigger-aware training objective that emphasizes target-relevant positions in poisoned samples, and theoretically prove that this objective is equivalent to training under an induced forward masking distribution. Unlike backdoors in autoregressive models, which typically manipulate next-token prediction, this characterization indicates that BadDLM can implant backdoors by exploiting the forward masking process. We instantiate BadDLM across different target levels: concept injection (BadDLM_Concept), semantic attribute steering (BadDLM_Attribute), alignment bypass (BadDLM_Align), and code payload injection (BadDLM_Payload). Experiments on mainstream open-source DLMs show that BadDLM achieves strong attack effectiveness across diverse targets while largely preserving benign utility, and remains effective against defenses designed for AR backdoors. Our findings expose a new class of security risks in diffusion-based language generation and call for defenses tailored to DLM denoising dynamics.

Xiaoyi Chen, Baisong Xin, Shengfang Zhai et al.

This paper finds that contrastive learning can produce superior sentence embeddings for pre-trained models but is also vulnerable to backdoor attacks. We present the first backdoor attack framework, BadCSE, for state-of-the-art sentence embeddings under supervised and unsupervised learning settings. The attack manipulates the construction of positive and negative pairs so that the backdoored samples have a similar embedding with the target sample (targeted attack) or the negative embedding of its clean version (non-targeted attack). By injecting the backdoor in sentence embeddings, BadCSE is resistant against downstream fine-tuning. We evaluate BadCSE on both STS tasks and other downstream tasks. The supervised non-targeted attack obtains a performance degradation of 194.86%, and the targeted attack maps the backdoored samples to the target embedding with a 97.70% success rate while maintaining the model utility.

Yue Liu, Hongcheng Gao, Shengfang Zhai et al. · tsinghua

As LLMs increasingly impact safety-critical applications, ensuring their safety using guardrails remains a key challenge. This paper proposes GuardReasoner, a new safeguard for LLMs, by guiding the guard model to learn to reason. Concretely, we first create the GuardReasonerTrain dataset, which consists of 127K samples with 460K detailed reasoning steps. Then, we introduce reasoning SFT to unlock the reasoning capability of guard models. In addition, we present hard sample DPO to further strengthen their reasoning ability. In this manner, GuardReasoner achieves better performance, explainability, and generalizability. Extensive experiments and analyses on 13 benchmarks of 3 guardrail tasks demonstrate its superiority. Remarkably, GuardReasoner 8B surpasses GPT-4o+CoT by 5.74% and LLaMA Guard 3 8B by 20.84% F1 score on average. We release the training data, code, and models with different scales (1B, 3B, 8B) of GuardReasoner : https://github.com/yueliu1999/GuardReasoner/.

Wenjie Qu, Ming Xu, Peiran Wang et al.

This position paper argues that securing LLM agents requires first defining an end-to-end correctness property that specifies when an agent's execution faithfully reflects the user's intent. Modern LLM agents operate over an \emph{intent-to-execution pipeline}, where natural-language instructions are translated into concrete system operations such as tool calls, API requests, and code execution. While recent defenses have made progress in constraining how agents construct tool calls, most existing formulations implicitly assume that tools are trusted. The emergence of systems such as OpenClaw, with open ecosystems of third-party skills and direct access to user environments, breaks this assumption and exposes new failure modes, including malicious or over-privileged components in the execution pipeline. Despite rapid progress in defense mechanisms, there is no adequate correctness property that defines what ``secure'' means for LLM agents, nor a principled way to evaluate the coverage of existing defenses. We observe that LLM agents are structurally analogous to compilers, where security violations correspond to mis-executions that do not preserve user intent. Drawing on this analogy, we identify two fundamental problem sources -- untrusted data ingestion and untrusted tool execution -- and derive four integrity properties that must hold simultaneously: \emph{Tool Integrity}, \emph{Instruction Integrity}, \emph{Judgment Integrity}, and \emph{Data Flow Integrity}. We call their conjunction \emph{intent-to-execution integrity}. Analyzing existing agentic defenses against these properties reveals that current systems provide only partial and non-compositional coverage, leaving fundamental gaps in securing modern LLM agents.

Zhihua Tian, Sirun Nan, Ming Xu et al.

Text-to-image (T2I) diffusion models have achieved remarkable progress in generating high-quality images but also raise people's concerns about generating harmful or misleading content. While extensive approaches have been proposed to erase unwanted concepts without requiring retraining from scratch, they inadvertently degrade performance on normal generation tasks. In this work, we propose Interpret then Deactivate (ItD), a novel framework to enable precise concept removal in T2I diffusion models while preserving overall performance. ItD first employs a sparse autoencoder (SAE) to interpret each concept as a combination of multiple features. By permanently deactivating the specific features associated with target concepts, we repurpose SAE as a zero-shot classifier that identifies whether the input prompt includes target concepts, allowing selective concept erasure in diffusion models. Moreover, we demonstrate that ItD can be easily extended to erase multiple concepts without requiring further training. Comprehensive experiments across celebrity identities, artistic styles, and explicit content demonstrate ItD's effectiveness in eliminating targeted concepts without interfering with normal concept generation. Additionally, ItD is also robust against adversarial prompts designed to circumvent content filters. Code is available at: https://github.com/NANSirun/Interpret-then-deactivate.

Shengfang Zhai, Weilong Wang, Jiajun Li et al.

Recently text-to-image models have gained widespread attention in the community due to their controllable and high-quality generation ability. However, the robustness of such models and their potential ethical issues have not been fully explored. In this paper, we introduce Universal Semantic Trigger, a meaningless token sequence that can be added at any location within the input text yet can induce generated images towards a preset semantic target.To thoroughly investigate it, we propose Semantic Gradient-based Search (SGS) framework. SGS automatically discovers the potential universal semantic triggers based on the given semantic targets. Furthermore, we design evaluation metrics to comprehensively evaluate semantic shift of images caused by these triggers. And our empirical analyses reveal that the mainstream open-source text-to-image models are vulnerable to our triggers, which could pose significant ethical threats. Our work contributes to a further understanding of text-to-image synthesis and helps users to automatically auditing their models before deployment.

Fanqing Meng, Lingxiao Du, Zijian Wu et al.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

Shengfang Zhai, Huanran Chen, Yinpeng Dong et al.

Text-to-image diffusion models have achieved tremendous success in the field of controllable image generation, while also coming along with issues of privacy leakage and data copyrights. Membership inference arises in these contexts as a potential auditing method for detecting unauthorized data usage. While some efforts have been made on diffusion models, they are not applicable to text-to-image diffusion models due to the high computation overhead and enhanced generalization capabilities. In this paper, we first identify a conditional overfitting phenomenon in text-to-image diffusion models, indicating that these models tend to overfit the conditional distribution of images given the corresponding text rather than the marginal distribution of images only. Based on this observation, we derive an analytical indicator, namely Conditional Likelihood Discrepancy (CLiD), to perform membership inference, which reduces the stochasticity in estimating memorization of individual samples. Experimental results demonstrate that our method significantly outperforms previous methods across various data distributions and dataset scales. Additionally, our method shows superior resistance to overfitting mitigation strategies, such as early stopping and data augmentation.

Luyuan Xie, Cong Li, Xin Zhang et al. · pku

Representation learning frameworks in unlabeled time series have been proposed for medical signal processing. Despite the numerous excellent progresses have been made in previous works, we observe the representation extracted for the time series still does not generalize well. In this paper, we present a Time series (medical signal) Representation Learning framework via Spectrogram (TRLS) to get more informative representations. We transform the input time-domain medical signals into spectrograms and design a time-frequency encoder named Time Frequency RNN (TFRNN) to capture more robust multi-scale representations from the augmented spectrograms. Our TRLS takes spectrogram as input with two types of different data augmentations and maximizes the similarity between positive ones, which effectively circumvents the problem of designing negative samples. Our evaluation of four real-world medical signal datasets focusing on medical signal classification shows that TRLS is superior to the existing frameworks.

Yuhao Wang, Wenjie Qu, Shengfang Zhai et al.

Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by incorporating external knowledge bases, but this may expose them to extraction attacks, leading to potential copyright and privacy risks. However, existing extraction methods typically rely on malicious inputs such as prompt injection or jailbreaking, making them easily detectable via input- or output-level detection. In this paper, we introduce Implicit Knowledge Extraction Attack (IKEA), which conducts Knowledge Extraction on RAG systems through benign queries. Specifically, IKEA first leverages anchor concepts-keywords related to internal knowledge-to generate queries with a natural appearance, and then designs two mechanisms that lead anchor concepts to thoroughly "explore" the RAG's knowledge: (1) Experience Reflection Sampling, which samples anchor concepts based on past query-response histories, ensuring their relevance to the topic; (2) Trust Region Directed Mutation, which iteratively mutates anchor concepts under similarity constraints to further exploit the embedding space. Extensive experiments demonstrate IKEA's effectiveness under various defenses, surpassing baselines by over 80% in extraction efficiency and 90% in attack success rate. Moreover, the substitute RAG system built from IKEA's extractions shows comparable performance to the original RAG and outperforms those based on baselines across multiple evaluation tasks, underscoring the stealthy copyright infringement risk in RAG systems.

Qiqi Lin, Xiaoyang Ji, Shengfang Zhai et al.

Large language models (LLMs) have achieved remarkable success in natural language processing, yet their performance and computational costs vary significantly. LLM routers play a crucial role in dynamically balancing these trade-offs. While previous studies have primarily focused on routing efficiency, security vulnerabilities throughout the entire LLM router life cycle, from training to inference, remain largely unexplored. In this paper, we present a comprehensive investigation into the life-cycle routing vulnerabilities of LLM routers. We evaluate both white-box and black-box adversarial robustness, as well as backdoor robustness, across several representative routing models under extensive experimental settings. Our experiments uncover several key findings: 1) Mainstream DNN-based routers tend to exhibit the weakest adversarial and backdoor robustness, largely due to their strong feature extraction capabilities that amplify vulnerabilities during both training and inference; 2) Training-free routers demonstrate the strongest robustness across different attack types, benefiting from the absence of learnable parameters that can be manipulated. These findings highlight critical security risks spanning the entire life cycle of LLM routers and provide insights for developing more robust models.

Qichen Zhao, Shengfang Zhai, Xinjian Bai et al.

Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.

Linyu Wu, Linhao Zhong, Wenjie Qu et al.

Diffusion large language models (dLLMs) offer faster generation than autoregressive models while maintaining comparable quality, but existing watermarking methods fail on them due to their non-sequential decoding. Unlike autoregressive models that generate tokens left-to-right, dLLMs can finalize tokens in arbitrary order, breaking the causal design underlying traditional watermarks. We present DMark, the first watermarking framework designed specifically for dLLMs. DMark introduces three complementary strategies to restore watermark detectability: predictive watermarking uses model-predicted tokens when actual context is unavailable; bidirectional watermarking exploits both forward and backward dependencies unique to diffusion decoding; and predictive-bidirectional watermarking combines both approaches to maximize detection strength. Experiments across multiple dLLMs show that DMark achieves 92.0-99.5% detection rates at 1% false positive rate while maintaining text quality, compared to only 49.6-71.2% for naive adaptations of existing methods. DMark also demonstrates robustness against text manipulations, establishing that effective watermarking is feasible for non-autoregressive language models.

Shengfang Zhai, Yinpeng Dong, Qingni Shen et al.

With the help of conditioning mechanisms, the state-of-the-art diffusion models have achieved tremendous success in guided image generation, particularly in text-to-image synthesis. To gain a better understanding of the training process and potential risks of text-to-image synthesis, we perform a systematic investigation of backdoor attack on text-to-image diffusion models and propose BadT2I, a general multimodal backdoor attack framework that tampers with image synthesis in diverse semantic levels. Specifically, we perform backdoor attacks on three levels of the vision semantics: Pixel-Backdoor, Object-Backdoor and Style-Backdoor. By utilizing a regularization loss, our methods efficiently inject backdoors into a large-scale text-to-image diffusion model while preserving its utility with benign inputs. We conduct empirical experiments on Stable Diffusion, the widely-used text-to-image diffusion model, demonstrating that the large-scale diffusion model can be easily backdoored within a few fine-tuning steps. We conduct additional experiments to explore the impact of different types of textual triggers, as well as the backdoor persistence during further training, providing insights for the development of backdoor defense methods. Besides, our investigation may contribute to the copyright protection of text-to-image models in the future.