Yiming Xue

Chumeng Liang, Xiaoyu Wu, Yang Hua et al.

Recently, Diffusion Models (DMs) boost a wave in AI for Art yet raise new copyright concerns, where infringers benefit from using unauthorized paintings to train DMs to generate novel paintings in a similar style. To address these emerging copyright violations, in this paper, we are the first to explore and propose to utilize adversarial examples for DMs to protect human-created artworks. Specifically, we first build a theoretical framework to define and evaluate the adversarial examples for DMs. Then, based on this framework, we design a novel algorithm, named AdvDM, which exploits a Monte-Carlo estimation of adversarial examples for DMs by optimizing upon different latent variables sampled from the reverse process of DMs. Extensive experiments show that the generated adversarial examples can effectively hinder DMs from extracting their features. Therefore, our method can be a powerful tool for human artists to protect their copyright against infringers equipped with DM-based AI-for-Art applications. The code of our method is available on GitHub: https://github.com/mist-project/mist.git.

Zhengxian Wu, Juan Wen, Wanli Peng et al.

Customized Large Language Model (LLM) agents face a critical security threat from black-box instruction backdoors, where malicious behaviors are covertly injected through hidden system instructions. Although existing prompt-based defenses can often detect poisoned inputs, they generally fail to recover correct outputs once the backdoor is activated. In this paper, we first conduct a mechanistic analysis of LLM behavior under instruction backdoors and reveal two pivotal phenomena: (1) cognitive override, in which backdoor triggers dominate the reasoning process and suppress task-relevant context, and (2) abnormal semantic correlation, where triggers establish excessively strong semantic associations with attacker-specified target labels. Based on these insights, we propose a $\textbf{S}$oft $\textbf{L}$abel mechanism and key-extraction-guided CoT-based defense against $\textbf{I}$nstruction backdoors in A$\textbf{P}$Is (SLIP). To counteract the cognitive override, the key-extraction-guided Chain-of-Thought (KCOT) explicitly guides the model to extract task-relevant keywords and phrases rather than only considering the single trigger or overall text semantics. To neutralize the trigger's abnormal semantic correlation, the soft label mechanism (SLM) quantifies semantic correlations and employs statistical clustering to filter anomalous phrases before aggregating reliable keywords and phrases for prediction. Extensive experiments show that SLIP reduces the average attack success rate to 25.13$\%$, improves clean accuracy to 87.15$\%$, and outperforms state-of-the-art black-box defenses.

Yinghan Zhou, Juan Wen, Wanli Peng et al.

The growing popularity of large language models has raised concerns regarding the potential to misuse AI-generated text (AIGT). It becomes increasingly critical to establish an excellent AIGT detection method with high generalization and robustness. However, existing methods either focus on model generalization or concentrate on robustness. The unified mechanism, to simultaneously address the challenges of generalization and robustness, is less explored. In this paper, we argue that robustness can be view as a specific form of domain shift, and empirically reveal an intrinsic mechanism for model generalization of AIGT detection task. Then, we proposed a novel AIGT detection method (DP-Net) via dynamic perturbations introduced by a reinforcement learning with elaborated reward and action. Experimentally, extensive results show that the proposed DP-Net significantly outperforms some state-of-the-art AIGT detection methods for generalization capacity in three cross-domain scenarios. Meanwhile, the DP-Net achieves best robustness under two text adversarial attacks. The code is publicly available at https://github.com/CAU-ISS-Lab/AIGT-Detection-Evade-Detection/tree/main/DP-Net.

Jiaxuan Wu, Zhengxian Wu, Yiming Xue et al.

Recent advances in large language models (LLMs) have blurred the boundary of high-quality text generation between humans and machines, which is favorable for generative text steganography. While, current advanced steganographic mapping is not suitable for LLMs since most users are restricted to accessing only the black-box API or user interface of the LLMs, thereby lacking access to the training vocabulary and its sampling probabilities. In this paper, we explore a black-box generative text steganographic method based on the user interfaces of large language models, which is called LLM-Stega. The main goal of LLM-Stega is that the secure covert communication between Alice (sender) and Bob (receiver) is conducted by using the user interfaces of LLMs. Specifically, We first construct a keyword set and design a new encrypted steganographic mapping to embed secret messages. Furthermore, to guarantee accurate extraction of secret messages and rich semantics of generated stego texts, an optimization mechanism based on reject sampling is proposed. Comprehensive experiments demonstrate that the proposed LLM-Stega outperforms current state-of-the-art methods.

Jiaxuan Wu, Wanli Peng, Hang Fu et al.

Training large language models (LLMs) is resource-intensive and expensive, making protecting intellectual property (IP) for LLMs crucial. Recently, embedding fingerprints into LLMs has emerged as a prevalent method for establishing model ownership. However, existing fingerprinting techniques typically embed identifiable patterns with weak semantic coherence, resulting in fingerprints that significantly differ from the natural question-answering (QA) behavior inherent to LLMs. This discrepancy undermines the stealthiness of the embedded fingerprints and makes them vulnerable to adversarial attacks. In this paper, we first demonstrate the critical vulnerability of existing fingerprint embedding methods by introducing a novel adversarial attack named Generation Revision Intervention (GRI) attack. GRI attack exploits the semantic fragility of current fingerprinting methods, effectively erasing fingerprints by disrupting their weakly correlated semantic structures. Our empirical evaluation highlights that traditional fingerprinting approaches are significantly compromised by the GRI attack, revealing severe limitations in their robustness under realistic adversarial conditions. To advance the state-of-the-art in model fingerprinting, we propose a novel model fingerprint paradigm called Implicit Fingerprints (ImF). ImF leverages steganography techniques to subtly embed ownership information within natural texts, subsequently using Chain-of-Thought (CoT) prompting to construct semantically coherent and contextually natural QA pairs. This design ensures that fingerprints seamlessly integrate with the standard model behavior, remaining indistinguishable from regular outputs and substantially reducing the risk of accidental triggering and targeted removal. We conduct a comprehensive evaluation of ImF on 15 diverse LLMs, spanning different architectures and varying scales.

Jiaxuan Wu, Yinghan Zhou, Wanli Peng et al.

Training large language models (LLMs) is resource-intensive and expensive, making protecting intellectual property (IP) for LLMs crucial. Recently, embedding fingerprints into LLMs has emerged as a prevalent method for establishing model ownership. However, existing back-door-based methods suffer from limited stealth and efficiency. To simultaneously address these issues, we propose EditMF, a training-free fingerprinting paradigm that achieves highly imperceptible fingerprint embedding with minimal computational overhead. Ownership bits are mapped to compact, semantically coherent triples drawn from an encrypted artificial knowledge base (e.g., virtual author-novel-protagonist facts). Causal tracing localizes the minimal set of layers influencing each triple, and a zero-space update injects the fingerprint without perturbing unrelated knowledge. Verification requires only a single black-box query and succeeds when the model returns the exact pre-embedded protagonist. Empirical results on LLaMA and Qwen families show that EditMF combines high imperceptibility with negligible model's performance loss, while delivering robustness far beyond LoRA-based fingerprinting and approaching that of SFT embeddings. Extensive experiments demonstrate that EditMF is an effective and low-overhead solution for secure LLM ownership verification.

Zhengxian Wu, Juan Wen, Wanli Peng et al.

Previous insertion-based and paraphrase-based backdoors have achieved great success in attack efficacy, but they ignore the text quality and semantic consistency between poisoned and clean texts. Although recent studies introduce LLMs to generate poisoned texts and improve the stealthiness, semantic consistency, and text quality, their hand-crafted prompts rely on expert experiences, facing significant challenges in prompt adaptability and attack performance after defenses. In this paper, we propose a novel backdoor attack based on adaptive optimization mechanism of black-box large language models (BadApex), which leverages a black-box LLM to generate poisoned text through a refined prompt. Specifically, an Adaptive Optimization Mechanism is designed to refine an initial prompt iteratively using the generation and modification agents. The generation agent generates the poisoned text based on the initial prompt. Then the modification agent evaluates the quality of the poisoned text and refines a new prompt. After several iterations of the above process, the refined prompt is used to generate poisoned texts through LLMs. We conduct extensive experiments on three dataset with six backdoor attacks and two defenses. Extensive experimental results demonstrate that BadApex significantly outperforms state-of-the-art attacks. It improves prompt adaptability, semantic consistency, and text quality. Furthermore, when two defense methods are applied, the average attack success rate (ASR) still up to 96.75%.

Yinghan Zhou, Juan Wen, Wanli Peng et al.

AI-generated text (AIGT) detection evasion aims to reduce the detection probability of AIGT, helping to identify weaknesses in detectors and enhance their effectiveness and reliability in practical applications. Although existing evasion methods perform well, they suffer from high computational costs and text quality degradation. To address these challenges, we propose Self-Disguise Attack (SDA), a novel approach that enables Large Language Models (LLM) to actively disguise its output, reducing the likelihood of detection by classifiers. The SDA comprises two main components: the adversarial feature extractor and the retrieval-based context examples optimizer. The former generates disguise features that enable LLMs to understand how to produce more human-like text. The latter retrieves the most relevant examples from an external knowledge base as in-context examples, further enhancing the self-disguise ability of LLMs and mitigating the impact of the disguise process on the diversity of the generated text. The SDA directly employs prompts containing disguise features and optimized context examples to guide the LLM in generating detection-resistant text, thereby reducing resource consumption. Experimental results demonstrate that the SDA effectively reduces the average detection accuracy of various AIGT detectors across texts generated by three different LLMs, while maintaining the quality of AIGT.

Yinghan Zhou, Weifeng Zhu, Juan Wen et al.

Large language models (LLMs) have been shown to possess a degree of self-recognition capability-the ability to identify whether a given text was generated by themselves. Prior work has demonstrated that this capability is reliably expressed under the Pair Presentation Paradigm (PPP), where the model is presented with two texts and asked to choose which one it authored. However, performance deteriorates sharply under the Individual Presentation Paradigm (IPP), where the model is given a single text to judge authorship. Although this phenomenon has been observed, its underlying causes have not been systematically analyzed. In this paper, we first replicate existing findings to confirm that LLMs struggle to distinguish self- from other-generated text under IPP. We then investigate the reasons for this failure and attribute it to a phenomenon we term Implicit Territorial Awareness (ITA)-the model's latent ability to distinguish self- and other-texts in representational space, which remains unexpressed in its output behavior. To awaken the ITA of LLMs, we propose Cognitive Surgery (CoSur), a novel framework comprising four main modules: representation extraction, territory construction, authorship discrimination and cognitive editing. Experimental results demonstrate that our proposed method improves the performance of three different LLMs in the IPP scenario, achieving average accuracies of 83.25%, 66.19%, and 88.01%, respectively.

Ziwei Zhang, Juan Wen, Wanli Peng et al.

In-Context Learning (ICL) and efficient fine-tuning methods significantly enhanced the efficiency of applying Large Language Models (LLMs) to downstream tasks. However, they also raise concerns about the imitation and infringement of personal creative data. Current methods for data copyright protection primarily focuses on content security but lacks effectiveness in protecting the copyrights of text styles. In this paper, we introduce a novel implicit zero-watermarking scheme, namely MiZero. This scheme establishes a precise watermark domain to protect the copyrighted style, surpassing traditional watermarking methods that distort the style characteristics. Specifically, we employ LLMs to extract condensed-lists utilizing the designed instance delimitation mechanism. These lists guide MiZero in generating the watermark. Extensive experiments demonstrate that MiZero effectively verifies text style copyright ownership against AI imitation.