Zi Liang

Zi Liang, Haibo Hu, Qingqing Ye et al.

The drastic increase of large language models' (LLMs) parameters has led to a new research direction of fine-tuning-free downstream customization by prompts, i.e., task descriptions. While these prompt-based services (e.g. OpenAI's GPTs) play an important role in many businesses, there has emerged growing concerns about the prompt leakage, which undermines the intellectual properties of these services and causes downstream attacks. In this paper, we analyze the underlying mechanism of prompt leakage, which we refer to as prompt memorization, and develop corresponding defending strategies. By exploring the scaling laws in prompt extraction, we analyze key attributes that influence prompt extraction, including model sizes, prompt lengths, as well as the types of prompts. Then we propose two hypotheses that explain how LLMs expose their prompts. The first is attributed to the perplexity, i.e. the familiarity of LLMs to texts, whereas the second is based on the straightforward token translation path in attention matrices. To defend against such threats, we investigate whether alignments can undermine the extraction of prompts. We find that current LLMs, even those with safety alignments like GPT-4, are highly vulnerable to prompt extraction attacks, even under the most straightforward user attacks. Therefore, we put forward several defense strategies with the inspiration of our findings, which achieve 83.8\% and 71.0\% drop in the prompt extraction rate for Llama2-7B and GPT-3.5, respectively. Source code is avaliable at https://github.com/liangzid/PromptExtractionEval.

Shuo Zhang, Junzhou Zhao, Pinghui Wang et al.

Multi-action dialog policy, which generates multiple atomic dialog actions per turn, has been widely applied in task-oriented dialog systems to provide expressive and efficient system responses. Existing policy models usually imitate action combinations from the labeled multi-action dialog examples. Due to data limitations, they generalize poorly toward unseen dialog flows. While reinforcement learning-based methods are proposed to incorporate the service ratings from real users and user simulators as external supervision signals, they suffer from sparse and less credible dialog-level rewards. To cope with this problem, we explore to improve multi-action dialog policy learning with explicit and implicit turn-level user feedback received for historical predictions (i.e., logged user feedback) that are cost-efficient to collect and faithful to real-world scenarios. The task is challenging since the logged user feedback provides only partial label feedback limited to the particular historical dialog actions predicted by the agent. To fully exploit such feedback information, we propose BanditMatch, which addresses the task from a feedback-enhanced semi-supervised learning perspective with a hybrid objective of semi-supervised learning and bandit learning. BanditMatch integrates pseudo-labeling methods to better explore the action space through constructing full label feedback. Extensive experiments show that our BanditMatch outperforms the state-of-the-art methods by generating more concise and informative responses. The source code and the appendix of this paper can be obtained from https://github.com/ShuoZhangXJTU/BanditMatch.

Zi Liang, Qingqing Ye, Yanyun Wang et al.

Model extraction attacks (MEAs) on large language models (LLMs) have received increasing attention in recent research. However, existing attack methods typically adapt the extraction strategies originally developed for deep neural networks (DNNs). They neglect the underlying inconsistency between the training tasks of MEA and LLM alignment, leading to suboptimal attack performance. To tackle this issue, we propose Locality Reinforced Distillation (LoRD), a novel model extraction algorithm specifically designed for LLMs. In particular, LoRD employs a newly defined policy-gradient-style training task that utilizes the responses of victim model as the signal to guide the crafting of preference for the local model. Theoretical analyses demonstrate that I) The convergence procedure of LoRD in model extraction is consistent with the alignment procedure of LLMs, and II) LoRD can reduce query complexity while mitigating watermark protection through our exploration-based stealing. Extensive experiments validate the superiority of our method in extracting various state-of-the-art commercial LLMs. Our code is available at: https://github.com/liangzid/LoRD-MEA .

Xiaoyu Xu, Minxin Du, Kun Fang et al.

Large language models (LLMs) demonstrate impressive capabilities across diverse tasks but raise concerns about privacy, copyright, and harmful materials. Existing LLM unlearning methods rarely consider the continual and high-volume nature of real-world deletion requests, which can cause utility degradation and catastrophic forgetting as requests accumulate. To address this challenge, we introduce \fit, a framework for continual unlearning that handles large numbers of deletion requests while maintaining robustness against both catastrophic forgetting and post-unlearning recovery. \fit mitigates degradation through rigorous data \underline{F}iltering, \underline{I}mportance-aware updates, and \underline{T}argeted layer attribution, enabling stable performance across long sequences of unlearning operations and achieving a favorable balance between forgetting effectiveness and utility retention. To support realistic evaluation, we present \textbf{PCH}, a benchmark covering \textbf{P}ersonal information, \textbf{C}opyright, and \textbf{H}armful content in sequential deletion scenarios, along with two symmetric metrics, Forget Degree (F.D.) and Retain Utility (R.U.), which jointly assess forgetting quality and utility preservation. Extensive experiments on four open-source LLMs with hundreds of deletion requests show that \fit achieves the strongest trade-off between F.D. and R.U., surpasses existing methods on MMLU, CommonsenseQA, and GSM8K, and remains resistant against both relearning and quantization recovery attacks.

Yanyun Wang, Yu Huang, Zi Liang et al.

The integration of audio modality into Large Audio Language Models (LALMs) significantly expands their attack surface. Existing jailbreak paradigms predominantly treat audio as a carrier for malicious payloads, relying on semantic optimization, acoustic parameter control, or additive perturbation to embed harmful content into the audio signal. In this work, we challenge this necessity and propose a new paradigm in which the role of audio shifts from content injection to safety alignment interference. We reveal that LALM safety alignment can be compromised solely by specific Acoustic Latent Semantics (ALS), the underlying paralinguistic features intrinsic to the priors of audio generative models. Distinct from previous works that leverage explicit acoustic parameters to merely style malicious audio, we demonstrate that interference audio, benign in content but infused with specific ALS, can serve as a universal jailbreak trigger. Leveraging this insight, we propose the Acoustic Interference Attack (AIA), which decouples the attack payload from the audio. Specifically, AIA employs a set of universal, instruction-neutral interference audio, enabling standard malicious text queries to bypass safety alignment without instance-specific optimization. Extensive experiments on 10 LALMs across five datasets demonstrate that AIA achieves the state-of-the-art attack success rate. Furthermore, our interpretability analysis uncovers the inference path drift induced by AIA and identifies the inherent effective patterns within ALS, revealing the fundamental vulnerability of cross-modal alignment in LALMs.

Yaxin Xiao, Qingqing Ye, Zi Liang et al.

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 at least 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 at least 70.15% in extracted models even under the combined MEA and WRK distortion, while preserving the utility of protected models.

Yanyun Wang, Dehui Du, Haibo Hu et al.

Recent years have witnessed the success of recurrent neural network (RNN) models in time series classification (TSC). However, neural networks (NNs) are vulnerable to adversarial samples, which cause real-life adversarial attacks that undermine the robustness of AI models. To date, most existing attacks target at feed-forward NNs and image recognition tasks, but they cannot perform well on RNN-based TSC. This is due to the cyclical computation of RNN, which prevents direct model differentiation. In addition, the high visual sensitivity of time series to perturbations also poses challenges to local objective optimization of adversarial samples. In this paper, we propose an efficient method called TSFool to craft highly-imperceptible adversarial time series for RNN-based TSC. The core idea is a new global optimization objective known as "Camouflage Coefficient" that captures the imperceptibility of adversarial samples from the class distribution. Based on this, we reduce the adversarial attack problem to a multi-objective optimization problem that enhances the perturbation quality. Furthermore, to speed up the optimization process, we propose to use a representation model for RNN to capture deeply embedded vulnerable samples whose features deviate from the latent manifold. Experiments on 11 UCR and UEA datasets showcase that TSFool significantly outperforms six white-box and three black-box benchmark attacks in terms of effectiveness, efficiency and imperceptibility from various perspectives including standard measure, human study and real-world defense.

Yanyun Wang, Li Liu, Zi Liang et al.

Adversarial Training (AT) is one of the most effective methods to enhance the robustness of Deep Neural Networks (DNNs). However, existing AT methods suffer from an inherent accuracy-robustness trade-off. Previous works have studied this issue under the current AT paradigm, but still face over 10% accuracy reduction without significant robustness improvement over simple baselines such as PGD-AT. This inherent trade-off raises a question: Whether the current AT paradigm, which assumes to learn corresponding benign and adversarial samples as the same class, inappropriately mixes clean and robust objectives that may be essentially inconsistent. In fact, our empirical results show that up to 40% of CIFAR-10 adversarial samples always fail to satisfy such an assumption across various AT methods and robust models, explicitly indicating the room for improvement of the current AT paradigm. To relax from this overstrict assumption and the tension between clean and robust learning, in this work, we propose a new AT paradigm by introducing an additional dummy class for each original class, aiming to accommodate hard adversarial samples with shifted distribution after perturbation. The robustness w.r.t. these adversarial samples can be achieved by runtime recovery from the predicted dummy classes to the corresponding original ones, without conflicting with the clean objective on accuracy of benign samples. Finally, based on our new paradigm, we propose a novel DUmmy Classes-based Adversarial Training (DUCAT) method that concurrently improves accuracy and robustness in a plug-and-play manner only relevant to logits, loss, and a proposed two-hot soft label-based supervised signal. Our method outperforms state-of-the-art (SOTA) benchmarks, effectively releasing the current trade-off. The code is available at https://github.com/FlaAI/DUCAT.

Zi Liang, Liantong Yu, Shiyu Zhang et al.

Overestimation in evaluating large language models (LLMs) has become an increasing concern. Due to the contamination of public benchmarks or imbalanced model training, LLMs may achieve unreal evaluation results on public benchmarks, either intentionally or unintentionally, which leads to unfair comparisons among LLMs and undermines their realistic capability assessments. Existing benchmarks attempt to address these issues by keeping test cases permanently secret, mitigating contamination through human evaluation, or repeatedly collecting and constructing new samples. However, these approaches fail to ensure reproducibility, transparency, and high efficiency simultaneously. Moreover, the extent of overestimation in current LLMs remains unquantified. To address these issues, we propose ArxivRoll, a dynamic evaluation framework inspired by one-time pad encryption in cryptography. ArxivRoll comprises two key components: \emph{i) SCP (Sequencing, Cloze, and Prediction)}, an automated generator for private test cases, and \emph{ii) Rugged Scores (RS)}, metrics that measure the proportion of public benchmark contamination and training bias. Leveraging SCP, ArxivRoll constructs a new benchmark every six months using recent articles from ArXiv and employs them for one-time evaluations of LLM performance. Extensive experiments demonstrate the high quality of our benchmark, and we provide a systematic evaluation of current LLMs. The source code is available at https://github.com/liangzid/ArxivRoll/.

Zi Liang, Ronghua Li, Yanyun Wang et al.

Large Language Model (LLM) agents have emerged as key intermediaries, orchestrating complex interactions between human users and a wide range of digital services and LLM infrastructures. While prior research has extensively examined the security of LLMs and agents in isolation, the systemic risk of the agent acting as a disruptive hub within the user-agent-service chain remains largely overlooked. In this work, we expose a novel threat paradigm by introducing Mobius Injection, a sophisticated attack that weaponizes autonomous agents into zombie nodes to launch what we define as gent-based and -Oriented DDoS (AbO-DDoS) attacks. By exploiting a structural vulnerability in agentic logic named Semantic Closure, an adversary can induce sustained recursive execution of agent components through a single textual injection. We demonstrate that this attack is exceptionally lightweight, stealthy against both traditional DDoS monitors and contemporary AI safety filters, and highly configurable, allowing for surgical targeting of specific environments or model providers. To evaluate the real-world impact, we conduct extensive experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs. Our results demonstrate that Mobius Injection achieves substantial attack success across diverse tasks, driving single-node call amplification up to 51.0x and multi-node p95 latency inflation up to 229.1x. The attack performance exhibits a superlinear increase with the number of poisoning nodes. To mitigate Mobius Injection, we propose a proactive defense mechanism using Agent Component Energy (ACE) Analysis, which detects malicious recursive triggers by measuring anomalous energy in the agent's component graph.

Zi Liang, Qipeng Xie, Jun He et al.

Recent advancements in Large Language Models (LLMs) have sparked interest in their application to Static Application Security Testing (SAST), primarily due to their superior contextual reasoning capabilities compared to traditional symbolic or rule-based methods. However, existing LLM-based approaches typically attempt to replace human experts directly without integrating effectively with existing SAST tools. This lack of integration results in ineffectiveness, including high rates of false positives, hallucinations, limited reasoning depth, and excessive token usage, making them impractical for industrial deployment. To overcome these limitations, we present a paradigm shift that reorchestrates the SAST workflow from current LLM-assisted structure to a new LLM-centered workflow. We introduce Argus (Agentic and Retrieval-Augmented Guarding System), the first multi-agent framework designed specifically for vulnerability detection. Argus incorporates three key novelties: comprehensive supply chain analysis, collaborative multi-agent workflows, and the integration of state-of-the-art techniques such as Retrieval-Augmented Generation (RAG) and ReAct to minimize hallucinations and enhance reasoning. Extensive empirical evaluation demonstrates that Argus significantly outperforms existing methods by detecting a higher volume of true vulnerabilities while simultaneously reducing false positives and operational costs. Notably, Argus has identified several critical zero-day vulnerabilities with CVE assignments.

Ao Ding, Hongzong Li, Zi Liang et al.

Large language models (LLMs) are increasingly deployed on edge devices under strict computation and quantization constraints, yet their security implications remain unclear. We study query-based knowledge extraction from quantized edge-deployed LLMs under realistic query budgets and show that, although quantization introduces noise, it does not remove the underlying semantic knowledge, allowing substantial behavioral recovery through carefully designed queries. To systematically analyze this risk, we propose \textbf{CLIQ} (\textbf{Cl}ustered \textbf{I}nstruction \textbf{Q}uerying), a structured query construction framework that improves semantic coverage while reducing redundancy. Experiments on quantized Qwen models (INT8/INT4) demonstrate that CLIQ consistently outperforms original queries across BERTScore, BLEU, and ROUGE, enabling more efficient extraction under limited budgets. These results indicate that quantization alone does not provide effective protection against query-based extraction, highlighting a previously underexplored security risk in edge-deployed LLMs.

Yanyun Wang, Qingqing Ye, Li Liu et al.

Adversarial Training (AT) is one of the most effective methods for developing robust deep neural networks (DNNs). However, AT faces a trade-off problem between clean accuracy and adversarial robustness. In this work, we reveal a surprising phenomenon for the first time: Varying input perturbation intensities for training samples near decision boundaries in AT have minimal impact on model robustness. This finding directly exposes the inconsistency between accuracy and robustness score fluctuations, leading us to identify the misalignment between input and latent spaces as a critical driver of the robustness-accuracy trade-off. To mitigate this misalignment for harmonizing accuracy and robustness, we define Robust Alignment as a new AT target, encouraging the model perception to change with input perturbations provided the final label prediction remains unchanged, which can be achieved via two novel ideas. First, we suggest a reduced and fixed perturbation intensity for those boundary samples, which facilitates the model to utilize the perturbations as learnable patterns, instead of noises that complicate decision boundaries meaninglessly. Second, we propose a Domain Interpolation Consistency Adversarial Regularization (DICAR), based on rigorous theoretical derivations, which explicitly introduces semantic alignment between input and latent spaces into AT. Based on these two ideas, we end up with a new Robust Alignment Adversarial Training (RAAT) method, effectively harmonizing accuracy and robustness. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-28-10 demonstrate the effectiveness of RAAT in improving the trade-off beyond four common baselines and a total of 14 related state-of-the-art (SOTA) works.

Zi Liang, Haibo Hu, Qingqing Ye et al.

Low rank adaptation (LoRA) has emerged as a prominent technique for fine-tuning large language models (LLMs) thanks to its superb efficiency gains over previous methods. While extensive studies have examined the performance and structural properties of LoRA, its behavior upon training-time attacks remain underexplored, posing significant security risks. In this paper, we theoretically investigate the security implications of LoRA's low-rank structure during fine-tuning, in the context of its robustness against data poisoning and backdoor attacks. We propose an analytical framework that models LoRA's training dynamics, employs the neural tangent kernel to simplify the analysis of the training process, and applies information theory to establish connections between LoRA's low rank structure and its vulnerability against training-time attacks. Our analysis indicates that LoRA exhibits better robustness to backdoor attacks than full fine-tuning, while becomes more vulnerable to untargeted data poisoning due to its over-simplified information geometry. Extensive experimental evaluations have corroborated our theoretical findings.

Xiaoyu Xu, Minxin Du, Zitong Li et al.

Although machine unlearning is essential for removing private, harmful, or copyrighted content from LLMs, current benchmarks often fail to faithfully represent the true "forgetting scope" learned by the model. We formalize two distinct unlearning granularities, domain-level and instance-level, and propose BiForget, an automated framework for synthesizing high-quality forget sets. Unlike prior work relying on external generators, BiForget exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. Our experiments across diverse benchmarks show that it achieves a superior balance of relevance, diversity, and efficiency. Quantitatively, in the Harry Potter domain, it improves relevance by ${\sim}20$ and diversity by ${\sim}$0.05 while halving the total data size compared to SOTAs. Ultimately, it facilitates more robust forgetting and better utility preservation, providing a more rigorous foundation for evaluating LLM unlearning.

Zi Liang, Zhiyao Wu, Haoyang Shang et al.

Decision boundary, the subspace of inputs where a machine learning model assigns equal classification probabilities to two classes, is pivotal in revealing core model properties and interpreting behaviors. While analyzing the decision boundary of large language models (LLMs) has raised increasing attention recently, constructing it for mainstream LLMs remains computationally infeasible due to the enormous vocabulary-sequence sizes and the auto-regressive nature of LLMs. To address this issue, in this paper we propose Decision Potential Surface (DPS), a new notion for analyzing LLM decision boundary. DPS is defined on the confidences in distinguishing different sampling sequences for each input, which naturally captures the potential of decision boundary. We prove that the zero-height isohypse in DPS is equivalent to the decision boundary of an LLM, with enclosed regions representing decision regions. By leveraging DPS, for the first time in the literature, we propose an approximate decision boundary construction algorithm, namely $K$-DPS, which only requires K-finite times of sequence sampling to approximate an LLM's decision boundary with negligible error. We theoretically derive the upper bounds for the absolute error, expected error, and the error concentration between K-DPS and the ideal DPS, demonstrating that such errors can be trade-off with sampling times. Our results are empirically validated by extensive experiments across various LLMs and corpora.

Zi Liang, Qingqing Ye, Xuan Liu et al.

Synthetic data refers to artificial samples generated by models. While it has been validated to significantly enhance the performance of large language models (LLMs) during training and has been widely adopted in LLM development, potential security risks it may introduce remain uninvestigated. This paper systematically evaluates the resilience of synthetic-data-integrated training paradigm for LLMs against mainstream poisoning and backdoor attacks. We reveal that such a paradigm exhibits strong resistance to existing attacks, primarily thanks to the different distribution patterns between poisoning data and queries used to generate synthetic samples. To enhance the effectiveness of these attacks and further investigate the security risks introduced by synthetic data, we introduce a novel and universal attack framework, namely, Virus Infection Attack (VIA), which enables the propagation of current attacks through synthetic data even under purely clean queries. Inspired by the principles of virus design in cybersecurity, VIA conceals the poisoning payload within a protective "shell" and strategically searches for optimal hijacking points in benign samples to maximize the likelihood of generating malicious content. Extensive experiments on both data poisoning and backdoor attacks show that VIA significantly increases the presence of poisoning content in synthetic data and correspondingly raises the attack success rate (ASR) on downstream models to levels comparable to those observed in the poisoned upstream models.

Yaxin Xiao, Qingqing Ye, Li Hu et al.

Machine unlearning enables the removal of specific data from ML models to uphold the right to be forgotten. While approximate unlearning algorithms offer efficient alternatives to full retraining, this work reveals that they fail to adequately protect the privacy of unlearned data. In particular, these algorithms introduce implicit residuals which facilitate privacy attacks targeting at unlearned data. We observe that these residuals persist regardless of model architectures, parameters, and unlearning algorithms, exposing a new attack surface beyond conventional output-based leakage. Based on this insight, we propose the Reminiscence Attack (ReA), which amplifies the correlation between residuals and membership privacy through targeted fine-tuning processes. ReA achieves up to 1.90x and 1.12x higher accuracy than prior attacks when inferring class-wise and sample-wise membership, respectively. To mitigate such residual-induced privacy risk, we develop a dual-phase approximate unlearning framework that first eliminates deep-layer unlearned data traces and then enforces convergence stability to prevent models from "pseudo-convergence", where their outputs are similar to retrained models but still preserve unlearned residuals. Our framework works for both classification and generation tasks. Experimental evaluations confirm that our approach maintains high unlearning efficacy, while reducing the adaptive privacy attack accuracy to nearly random guess, at the computational cost of 2-12% of full retraining from scratch.

Nuo Xu, Pinghui Wang, Zi Liang et al.

Legal case retrieval (LCR) aims to automatically scour for comparable legal cases based on a given query, which is crucial for offering relevant precedents to support the judgment in intelligent legal systems. Due to similar goals, it is often associated with a similar case matching (LCM) task. To address them, a daunting challenge is assessing the uniquely defined legal-rational similarity within the judicial domain, which distinctly deviates from the semantic similarities in general text retrieval. Past works either tagged domain-specific factors or incorporated reference laws to capture legal-rational information. However, their heavy reliance on expert or unrealistic assumptions restricts their practical applicability in real-world scenarios. In this paper, we propose an end-to-end model named LCM-LAI to solve the above challenges. Through meticulous theoretical analysis, LCM-LAI employs a dependent multi-task learning framework to capture legal-rational information within legal cases by a law article prediction (LAP) sub-task, without any additional assumptions in inference. Besides, LCM-LAI proposes an article-aware attention mechanism to evaluate the legal-rational similarity between across-case sentences based on law distribution, which is more effective than conventional semantic similarity. Weperform a series of exhaustive experiments including two different tasks involving four real-world datasets. Results demonstrate that LCM-LAI achieves state-of-the-art performance.

Zi Liang, Pinghui Wang, Ruofei Zhang et al.

The drastic increase in language models' parameters has led to a new trend of deploying models in cloud servers, raising growing concerns about private inference for Transformer-based models. Existing two-party privacy-preserving techniques, however, only take into account natural language understanding (NLU) scenarios. Private inference in natural language generation (NLG), crucial for applications like translation and code completion, remains underexplored.In addition, previous privacy-preserving techniques suffer from convergence issues during model training and exhibit poor inference speed when used with NLG models due to the neglect of time-consuming operations in auto-regressive generations. To address these issues, we propose a fast private text generation framework for Transformer-based language models, namely MERGE.MERGE reuses the output hidden state as the word embedding to bypass the embedding computation and reorganize the linear operations in the Transformer module to accelerate the forward procedure. Extensive experiments show that MERGE achieves a 26.5x speedup to the vanilla encrypted model under the sequence length 512, and reduces 80\% communication cost, with an up to 10x speedup to state-of-the-art approximated models.

Zi Liang, Pinghui Wang, Ruofei Zhang et al.

Recent years have seen increasing concerns about the unsafe response generation of large-scale dialogue systems, where agents will learn offensive or biased behaviors from the real-world corpus. Some methods are proposed to address the above issue by detecting and replacing unsafe training examples in a pipeline style. Though effective, they suffer from a high annotation cost and adapt poorly to unseen scenarios as well as adversarial attacks. Besides, the neglect of providing safe responses (e.g. simply replacing with templates) will cause the information-missing problem of dialogues. To address these issues, we propose an unsupervised pseudo-label sampling method, TEMP, that can automatically assign potential safe responses. Specifically, our TEMP method groups responses into several clusters and samples multiple labels with an adaptively sharpened sampling strategy, inspired by the observation that unsafe samples in the clusters are usually few and distribute in the tail. Extensive experiments in chitchat and task-oriented dialogues show that our TEMP outperforms state-of-the-art models with weak supervision signals and obtains comparable results under unsupervised learning settings.