Zeyang Sha

Zeyang Sha, Zheng Li, Ning Yu et al.

Text-to-image generation models that generate images based on prompt descriptions have attracted an increasing amount of attention during the past few months. Despite their encouraging performance, these models raise concerns about the misuse of their generated fake images. To tackle this problem, we pioneer a systematic study on the detection and attribution of fake images generated by text-to-image generation models. Concretely, we first build a machine learning classifier to detect the fake images generated by various text-to-image generation models. We then attribute these fake images to their source models, such that model owners can be held responsible for their models' misuse. We further investigate how prompts that generate fake images affect detection and attribution. We conduct extensive experiments on four popular text-to-image generation models, including DALL$\cdot$E 2, Stable Diffusion, GLIDE, and Latent Diffusion, and two benchmark prompt-image datasets. Empirical results show that (1) fake images generated by various models can be distinguished from real ones, as there exists a common artifact shared by fake images from different models; (2) fake images can be effectively attributed to their source models, as different models leave unique fingerprints in their generated images; (3) prompts with the ``person'' topic or a length between 25 and 75 enable models to generate fake images with higher authenticity. All findings contribute to the community's insight into the threats caused by text-to-image generation models. We appeal to the community's consideration of the counterpart solutions, like ours, against the rapidly-evolving fake image generation.

Zeyang Sha, Xinlei He, Pascal Berrang et al.

Backdoor attacks represent one of the major threats to machine learning models. Various efforts have been made to mitigate backdoors. However, existing defenses have become increasingly complex and often require high computational resources or may also jeopardize models' utility. In this work, we show that fine-tuning, one of the most common and easy-to-adopt machine learning training operations, can effectively remove backdoors from machine learning models while maintaining high model utility. Extensive experiments over three machine learning paradigms show that fine-tuning and our newly proposed super-fine-tuning achieve strong defense performance. Furthermore, we coin a new term, namely backdoor sequela, to measure the changes in model vulnerabilities to other attacks before and after the backdoor has been removed. Empirical evaluation shows that, compared to other defense methods, super-fine-tuning leaves limited backdoor sequela. We hope our results can help machine learning model owners better protect their models from backdoor threats. Also, it calls for the design of more advanced attacks in order to comprehensively assess machine learning models' backdoor vulnerabilities.

Ziqing Yang, Zeyang Sha, Michael Backes et al.

Visual prompt learning, as a newly emerged technique, leverages the knowledge learned by a large-scale pre-trained model and adapts it to downstream tasks through the usage of prompts. While previous research has focused on designing effective prompts, in this work, we argue that compared to prompt design, a good mapping strategy matters more. In this sense, we propose SeMap, a more effective mapping using the semantic alignment between the pre-trained model's knowledge and the downstream task. Our experimental results show that SeMap can largely boost the performance of visual prompt learning. Moreover, our experiments show that SeMap is capable of achieving competitive zero-shot transfer, indicating that it can perform the downstream task without any fine-tuning on the corresponding dataset. This demonstrates the potential of our proposed method to be used in a broader range of applications where the zero-shot transfer is desired. Results suggest that our proposed SeMap could lead to significant advancements in both visual prompt learning and zero-shot transfer. We hope with SeMap, we can help the community move forward to more efficient and lightweight utilization of large vision models.

Boyang Zhang, Xinyue Shen, Wai Man Si et al.

Moderating offensive, hateful, and toxic language has always been an important but challenging topic in the domain of safe use in NLP. The emerging large language models (LLMs), such as ChatGPT, can potentially further accentuate this threat. Previous works have discovered that ChatGPT can generate toxic responses using carefully crafted inputs. However, limited research has been done to systematically examine when ChatGPT generates toxic responses. In this paper, we comprehensively evaluate the toxicity in ChatGPT by utilizing instruction-tuning datasets that closely align with real-world scenarios. Our results show that ChatGPT's toxicity varies based on different properties and settings of the prompts, including tasks, domains, length, and languages. Notably, prompts in creative writing tasks can be 2x more likely than others to elicit toxic responses. Prompting in German and Portuguese can also double the response toxicity. Additionally, we discover that certain deliberately toxic prompts, designed in earlier studies, no longer yield harmful responses. We hope our discoveries can guide model developers to better regulate these AI systems and the users to avoid undesirable outputs.

Zeyang Sha, Yang Zhang

The increasing reliance on large language models (LLMs) such as ChatGPT in various fields emphasizes the importance of ``prompt engineering,'' a technology to improve the quality of model outputs. With companies investing significantly in expert prompt engineers and educational resources rising to meet market demand, designing high-quality prompts has become an intriguing challenge. In this paper, we propose a novel attack against LLMs, named prompt stealing attacks. Our proposed prompt stealing attack aims to steal these well-designed prompts based on the generated answers. The prompt stealing attack contains two primary modules: the parameter extractor and the prompt reconstruction. The goal of the parameter extractor is to figure out the properties of the original prompts. We first observe that most prompts fall into one of three categories: direct prompt, role-based prompt, and in-context prompt. Our parameter extractor first tries to distinguish the type of prompts based on the generated answers. Then, it can further predict which role or how many contexts are used based on the types of prompts. Following the parameter extractor, the prompt reconstructor can be used to reconstruct the original prompts based on the generated answers and the extracted features. The final goal of the prompt reconstructor is to generate the reversed prompts, which are similar to the original prompts. Our experimental results show the remarkable performance of our proposed attacks. Our proposed attacks add a new dimension to the study of prompt engineering and call for more attention to the security issues on LLMs.

Yule Liu, Jingyi Zheng, Zhen Sun et al.

Recent advancements in large reasoning models (LRMs) have demonstrated the effectiveness of scaling test-time computation to enhance reasoning capabilities on various tasks. However, LRMs often suffer from an ``overthinking'' problem, where the model generates excessively redundant reasoning steps with limited performance gains. In this work, we empirically reveal an important characteristic of LRM behaviors that placing external CoTs generated by smaller models between the thinking token (\texttt{<think>} and \texttt{</think>}) can effectively manipulate the model to generate fewer thoughts. Building on this finding, we propose a simple yet efficient pipeline, \Method, to enable LRMs to bypass unnecessary intermediate steps, thereby significantly reducing computational costs. We conduct extensive experiments to evaluate the utility and efficiency of \Method. For instance, when applied to QwQ-32B on the LiveBench/Code dataset, \Method keeps the original performance while reducing output token counts by approximately 30\%, with minimal overhead introduced by the CoT generator. Furthermore, we identify two suboptimal modes, blindly following flawed external thoughts and unnecessary rethinking, and show that simple mitigations, such as difficulty-aware fallbacks, can further improve performance. Overall, \Method offers a practical, general, and efficient way to optimize LRM inference, making powerful reasoning models more accessible and scalable for real-world applications.

Junjie Chu, Zeyang Sha, Michael Backes et al.

Significant advancements have recently been made in large language models represented by GPT models. Users frequently have multi-round private conversations with cloud-hosted GPT models for task optimization. Yet, this operational paradigm introduces additional attack surfaces, particularly in custom GPTs and hijacked chat sessions. In this paper, we introduce a straightforward yet potent Conversation Reconstruction Attack. This attack targets the contents of previous conversations between GPT models and benign users, i.e., the benign users' input contents during their interaction with GPT models. The adversary could induce GPT models to leak such contents by querying them with designed malicious prompts. Our comprehensive examination of privacy risks during the interactions with GPT models under this attack reveals GPT-4's considerable resilience. We present two advanced attacks targeting improved reconstruction of past conversations, demonstrating significant privacy leakage across all models under these advanced techniques. Evaluating various defense mechanisms, we find them ineffective against these attacks. Our findings highlight the ease with which privacy can be compromised in interactions with GPT models, urging the community to safeguard against potential abuses of these models' capabilities.

Zeyang Sha, Shiwen Cui, Weiqiang Wang

Recent advancements in Large Language Models(LLMs) have demonstrated their capabilities not only in reasoning but also in invoking external tools, particularly search engines. However, teaching models to discern when to invoke search and when to rely on their internal knowledge remains a significant challenge. Existing reinforcement learning approaches often lead to redundant search behaviors, resulting in inefficiencies and over-cost. In this paper, we propose SEM, a novel post-training reinforcement learning framework that explicitly trains LLMs to optimize search usage. By constructing a balanced dataset combining MuSiQue and MMLU, we create scenarios where the model must learn to distinguish between questions it can answer directly and those requiring external retrieval. We design a structured reasoning template and employ Group Relative Policy Optimization(GRPO) to post-train the model's search behaviors. Our reward function encourages accurate answering without unnecessary search while promoting effective retrieval when needed. Experimental results demonstrate that our method significantly reduces redundant search operations while maintaining or improving answer accuracy across multiple challenging benchmarks. This framework advances the model's reasoning efficiency and extends its capability to judiciously leverage external knowledge.

Zeyang Sha, Hanling Tian, Zhuoer Xu et al.

The emergence of autonomous Large Language Model (LLM) agents capable of tool usage has introduced new safety risks that go beyond traditional conversational misuse. These agents, empowered to execute external functions, are vulnerable to both user-initiated threats (e.g., adversarial prompts) and tool-initiated threats (e.g., malicious outputs from compromised tools). In this paper, we propose the first unified safety-alignment framework for tool-using agents, enabling models to handle both channels of threat via structured reasoning and sandboxed reinforcement learning. We introduce a tri-modal taxonomy, including benign, malicious, and sensitive for both user prompts and tool responses, and define a policy-driven decision model. Our framework employs a custom-designed sandbox environment that simulates real-world tool execution and allows fine-grained reward shaping. Through extensive evaluations on public and self-built benchmarks, including Agent SafetyBench, InjecAgent, and BFCL, we demonstrate that our safety-aligned agents significantly improve resistance to security threats while preserving strong utility on benign tasks. Our results show that safety and effectiveness can be jointly optimized, laying the groundwork for trustworthy deployment of autonomous LLM agents.

Zhen Sun, Ziyi Zhang, Zeren Luo et al.

Fine-grained edited image detection of localized edits in images is crucial for assessing content authenticity, especially given that modern diffusion models and image editing methods can produce highly realistic manipulations. However, this domain faces three challenges: (1) Binary classifiers yield only a global real-or-fake label without providing localization; (2) Traditional computer vision methods often rely on costly pixel-level annotations; and (3) No large-scale, high-quality dataset exists for modern image-editing detection techniques. To address these gaps, we develop an automated data-generation pipeline to create FragFake, the first dedicated benchmark dataset for edited image detection, which includes high-quality images from diverse editing models and a wide variety of edited objects. Based on FragFake, we utilize Vision Language Models (VLMs) for the first time in the task of edited image classification and edited region localization. Experimental results show that fine-tuned VLMs achieve higher average Object Precision across all datasets, significantly outperforming pretrained models. We further conduct ablation and transferability analyses to evaluate the detectors across various configurations and editing scenarios. To the best of our knowledge, this work is the first to reformulate localized image edit detection as a vision-language understanding task, establishing a new paradigm for the field. We anticipate that this work will establish a solid foundation to facilitate and inspire subsequent research endeavors in the domain of multimodal content authenticity.

Chenlong Yin, Zeyang Sha, Shiwen Cui et al.

Enhancing the reasoning capabilities of Large Language Models (LLMs) is a key strategy for building Agents that "think then act." However, recent observations, like OpenAI's o3, suggest a paradox: stronger reasoning often coincides with increased hallucination, yet no prior work has systematically examined whether reasoning enhancement itself causes tool hallucination. To address this gap, we pose the central question: Does strengthening reasoning increase tool hallucination? To answer this, we introduce SimpleToolHalluBench, a diagnostic benchmark measuring tool hallucination in two failure modes: (i) no tool available, and (ii) only distractor tools available. Through controlled experiments, we establish three key findings. First, we demonstrate a causal relationship: progressively enhancing reasoning through RL increases tool hallucination proportionally with task performance gains. Second, this effect transcends overfitting - training on non-tool tasks (e.g., mathematics) still amplifies subsequent tool hallucination. Third, the effect is method-agnostic, appearing when reasoning is instilled via supervised fine-tuning and when it is merely elicited at inference by switching from direct answers to step-by-step thinking. We also evaluate mitigation strategies including Prompt Engineering and Direct Preference Optimization (DPO), revealing a fundamental reliability-capability trade-off: reducing hallucination consistently degrades utility. Mechanistically, Reasoning RL disproportionately collapses tool-reliability-related representations, and hallucinations surface as amplified divergences concentrated in late-layer residual streams. These findings reveal that current reasoning enhancement methods inherently amplify tool hallucination, highlighting the need for new training objectives that jointly optimize for capability and reliability.

Zeyang Sha, Xinlei He, Ning Yu et al.

Self-supervised representation learning techniques have been developing rapidly to make full use of unlabeled images. They encode images into rich features that are oblivious to downstream tasks. Behind their revolutionary representation power, the requirements for dedicated model designs and a massive amount of computation resources expose image encoders to the risks of potential model stealing attacks - a cheap way to mimic the well-trained encoder performance while circumventing the demanding requirements. Yet conventional attacks only target supervised classifiers given their predicted labels and/or posteriors, which leaves the vulnerability of unsupervised encoders unexplored. In this paper, we first instantiate the conventional stealing attacks against encoders and demonstrate their severer vulnerability compared with downstream classifiers. To better leverage the rich representation of encoders, we further propose Cont-Steal, a contrastive-learning-based attack, and validate its improved stealing effectiveness in various experiment settings. As a takeaway, we appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours.