Large language models (LLMs) are powerful at question-answering but prone to hallucinations due to limited domain-specific or up-todate knowledge.Retrieval augmented generation (RAG) mitigates this by adding an external retriever and knowledge database, yet RAG remains vulnerable to targeted attacks that degrade outputs or manipulate opinions.Prior attacks typically assume adversaries know the service is RAG-enhanced and may even know deployment details, an assumption often invalid for real-world commercial LLMs that expose only black-box APIs.This opacity also risks misleading users about system capabilities.This work aims to bridge this gap by proposing RAG-ID, a framework for IDentifying RAG properties in LLM services.We classify adversaries into three knowledge levels and design six attack methods.Experiments show these attacks reliably detect RAG -up to 99.97% accuracy with partial or no optional knowledge, and nearly 100% when the LLM and database are known.After detection, RAG-ID can infer finer RAG properties (e.g., deployed LLM and knowledge database).We consider RAG-ID a reconnaissance tool for attackers, a way to facilitate users' transparent selection of LLM services, and a guide for RAG developers in refining security measures.

Machine learning (ML), driven by prominent paradigms such as centralized and federated learning, has made significant progress in various critical applications. However, its remarkable success has been accompanied by various attacks. Recently, the model hijacking attack has shown that ML models can be hijacked to execute tasks different from their original tasks, which increases both accountability and parasitic computational risks. Nevertheless, thus far, this attack has only focused on centralized learning. In this work, we broaden the scope of this attack to the federated learning domain, where multiple clients collaboratively train a global model without sharing their data. Specifically, we present the first-of-its-kind hijacking attack against the global model in federated learning, namely HijackFL. The adversary aims to force the global model to perform a different task (called hijacking task) from its original task without the server or benign client noticing. To accomplish this, unlike existing methods that use data poisoning to modify the target model’s parameters, HijackFL searches for pixel-level perturbations based on their local model (without modifications) to align hijacking samples with the original ones in the feature space. When performing the hijacking task, the adversary applies these perturbations to the hijacking samples, compelling the global model to identify them as original ones and predict them accordingly. Extensive experiments demonstrate HijackFL significantly outperforms baselines, e.g., 92.75% vs. 10%. We further investigate the factors that affect its performance and discuss possible defenses to mitigate its impact.

While Retrieval-Augmented Generation (RAG) effectively reduces hallucinations by integrating external knowledge bases, it introduces vulnerabilities to membership inference attacks (MIAs), particularly in systems handling sensitive data. Existing MIAs targeting RAG's external databases often rely on model responses but ignore the interference of non-member-retrieved documents on RAG outputs, limiting their effectiveness. To address this, we propose DCMI, a differential calibration MIA that mitigates the negative impact of non-member-retrieved documents. Specifically, DCMI leverages the sensitivity gap between member and non-member retrieved documents under query perturbation. It generates perturbed queries for calibration to isolate the contribution of member-retrieved documents while minimizing the interference from non-member-retrieved documents. Experiments under progressively relaxed assumptions show that DCMI consistently outperforms baselines—for example, achieving 97.42% AUC and 94.35% Accuracy against the RAG system with Flan-T5, exceeding the MBA baseline by over 40%. Furthermore, on real-world RAG platforms such as Dify and MaxKB, DCMI maintains a 10%-20% advantage over the baseline. These results highlight significant privacy risks in RAG systems and emphasize the need for stronger protection mechanisms. We appeal to the community's consideration of deeper investigations, like ours, against the data leakage risks in rapidly evolving RAG systems.

The open-source release of large language models (LLMs) enables malicious users to create unauthorized derivative models at low cost, posing significant threats to intellectual property (IP) and market stability. Existing IP protection methods either require access to model parameters or are vulnerable to fine-tuning attacks. To fill this gap, we propose ErrorTrace, a robust and black-box traceability mechanism for protecting LLM IP. Specifically, ErrorTrace leverages the unique error patterns of model families by mapping and analyzing their distinct error spaces, enabling robust and efficient IP protection without relying on internal parameters or specific query responses. Experimental results show that ErrorTrace achieves a traceability accuracy of 0.8518 for 27 base models when the suspect model is not included in ErrorTrace's training set, outperforming the baseline by 0.2593. Additionally,ErrorTrace successfully tracks 34 fine-tuned, pruned and merged models across various scenarios, demonstrating its broad applicability and robustness. In addition, ErrorTrace shows a certain level of resilience when subjected to adversarial attacks. Our code is available at: https://github.com/csdatazcc/ErrorTrace.

Text-to-image (T2I) generative models have revolutionized content creation by transforming textual descriptions into high-quality images. However, these models are vulnerable to jailbreaking attacks, where carefully crafted prompts bypass safety mechanisms to produce unsafe content. While researchers have developed various jailbreak attacks to expose this risk, these methods face significant limitations, including impractical access requirements, easily detectable unnatural prompts, restricted search spaces, and high query demands on the target system. In this paper, we propose JailFuzzer, a novel fuzzing framework driven by large language model (LLM) agents, designed to efficiently generate natural and semantically meaningful jailbreak prompts in a black-box setting. Specifically, JailFuzzer employs fuzz-testing principles with three components: a seed pool for initial and jailbreak prompts, a guided mutation engine for generating meaningful variations, and an oracle function to evaluate jailbreak success. Furthermore, we construct the guided mutation engine and oracle function by LLM-based agents, which further ensures efficiency and adaptability in black-box settings. Extensive experiments demonstrate that JailFuzzer has significant advantages in jailbreaking T2I models. It generates natural and semantically coherent prompts, reducing the likelihood of detection by traditional defenses. Additionally, it achieves a high success rate in jailbreak attacks with minimal query overhead, outperforming existing methods across all key metrics. This study underscores the need for stronger safety mechanisms in generative models and provides a foundation for future research on defending against sophisticated jailbreaking attacks. JailFuzzer is open-source and available at this repository: https://github.com/YingkaiD/JailFuzzer.

Training a machine learning model with data following a meaningful order, i.e., from easy to hard, has been proven to be effective in accelerating the training process and achieving better model performance. The key enabling technique is curriculum learning (CL), which has seen great success and has been deployed in areas like image and text classification. Yet, how CL affects the privacy of machine learning is unclear. Given that CL changes the way a model memorizes the training data, its influence on data privacy needs to be thoroughly evaluated. To fill this knowledge gap, we perform the first study and leverage membership inference attack (MIA) and attribute inference attack (AIA) as two vectors to quantify the privacy leakage caused by CL. Our evaluation of 9 real-world datasets with attack methods (NN-based, metric-based, label-only MIA, and NN-based AIA) revealed new insights about CL. First, MIA becomes slightly more effective when CL is applied, but the impact is much more prominent to a subset of training samples ranked as difficult. Second, a model trained under CL is less vulnerable under AIA, compared to MIA. Third, the existing defense techniques like MemGuard and MixupMMD are not effective under CL. Finally, based on our insights into CL, we propose a new MIA, termed Diff-Cali, which exploits the difficulty scores for result calibration and is demonstrated to be effective against all CL methods and the normal training method. With this study, we hope to draw the community's attention to the unintended privacy risks of emerging machine-learning techniques and develop new attack benchmarks and defense solutions.

Machine Learning as a Service (MLaaS) platforms have gained popularity due to their accessibility, cost-efficiency, scalability, and rapid development capabilities. However, recent research has highlighted the vulnerability of cloud-based models in MLaaS to model extraction attacks. In this paper, we introduce FDINet, a novel defense mechanism that leverages the feature distribution of deep neural network (DNN) models. Concretely, by analyzing the feature distribution from the adversary's queries, we reveal that the feature distribution of these queries deviates from that of the model's problem domain. Based on this key observation, we propose Feature Distortion Index (FDI), a metric designed to quantitatively measure the feature distribution deviation of received queries. The proposed FDINet utilizes FDI to train a binary detector and exploits FDI similarity to identify colluding adversaries from distributed extraction attacks. We conduct extensive experiments to evaluate FDINet against six state-of-the-art extraction attacks on four benchmark datasets and four popular model architectures. Empirical results demonstrate the following findings: (1) FDINet proves to be highly effective in detecting model extraction, achieving a100% detection accuracyon DFME and DaST. (2) FDINet is highly efficient, using just 50 queries to raise an extraction alarm with anaverage confidence of 96.08%for GTSRB. (3) FDINet exhibits the capability to identify colluding adversaries with an accuracyexceeding 91%. Additionally, it demonstrates the ability to detect two types of adaptive attacks.

The prediction module, powered by deep learning models, constitutes a fundamental component of high-level Autonomous Vehicles (AVs). Given the direct influence of the module’s prediction accuracy on AV driving behavior, ensuring its security is paramount. However, limited studies have explored the adversarial robustness of the prediction modules. Furthermore, existing methods still generate adversarial trajectories that deviate significantly from human driving behavior. These deviations can be easily identified as hazardous by AVs’ anomaly detection models and thus cannot effectively evaluate and reflect the robustness of the prediction modules. To bridge this gap, we propose a stealthy and more effective optimization-based attack method. Specifically, we reformulate the optimization problem using Lagrangian relaxation and design a Frenet-based objective function along with a distinct constraint space. We conduct extensive evaluations on 2 popular prediction models and 2 benchmark datasets. Our results show that our attack is highly effective, with over 87% attack success rates, outperforming all baseline attacks. Moreover, our attack method significantly improves the stealthiness of adversarial trajectories while guaranteeing adherence to physical constraints. Our attack is also found robust to noise from upstream modules, transferable across trajectory prediction models, and high realizability. Lastly, to verify its effectiveness in real-world applications, we conduct further simulation evaluations using a production-grade simulator. These simulations reveal that the adversarial trajectory we created could convincingly induce autonomous vehicles (AVs) to initiate hard braking.

Most existing membership inference attacks (MIAs) utilize metrics (e.g., loss) calculated on the model's final state, while recent advanced attacks leverage metrics computed at various stages, including both intermediate and final stages, throughout the model training. Nevertheless, these attacks often process multiple intermediate states of the metric independently, ignoring their time-dependent patterns. Consequently, they struggle to effectively distinguish between members and non-members who exhibit similar metric values, particularly resulting in a high false-positive rate.

Adapting Large Language Models (LLMs) to specific tasks introduces concerns about computational efficiency, prompting an exploration of efficient methods such as In-Context Learning (ICL). However, the vulnerability of ICL to privacy attacks under realistic assumptions remains largely unexplored. In this work, we present the first membership inference attack tailored for ICL, relying solely on generated texts without their associated probabilities. We propose four attack strategies tailored to various constrained scenarios and conduct extensive experiments on four popular large language models. Empirical results show that our attacks can accurately determine membership status in most cases, e.g., 95% accuracy advantage against LLaMA, indicating that the associated risks are much higher than those shown by existing probability-based attacks. Additionally, we propose a hybrid attack that synthesizes the strengths of the aforementioned strategies, achieving an accuracy advantage of over 95% in most cases. Furthermore, we investigate three potential defenses targeting data, instruction, and output. Results demonstrate combining defenses from orthogonal dimensions significantly reduces privacy leakage and offers enhanced privacy assurances.

Fine-tuning pre-trained models for downstream tasks has led to a proliferation of open-sourced task-specific models. Recently, Model Merging (MM) has emerged as an effective approach to facilitate knowledge transfer among these independently fine-tuned models. MM directly combines multiple fine-tuned task-specific models into a merged model without additional training, and the resulting model shows enhanced capabilities in multiple tasks. Although MM provides great utility, it may come with security risks because an adversary can exploit MM to affect multiple downstream tasks. However, the security risks of MM have barely been studied. In this paper, we first find that MM, as a new learning paradigm, introduces unique challenges for existing backdoor attacks due to the merging process. To address these challenges, we introduce BadMerging, the first backdoor attack specifically designed for MM. Notably, BadMerging allows an adversary to compromise the entire merged model by contributing as few as one backdoored task-specific model. BadMerging comprises a two-stage attack mechanism and a novel feature-interpolation-based loss to enhance the robustness of embedded backdoors against the changes of different merging parameters. Considering that a merged model may incorporate tasks from different domains, BadMerging can jointly compromise the tasks provided by the adversary (on-task attack) and other contributors (off-task attack) and solve the corresponding unique challenges with novel attack designs. Extensive experiments show that BadMerging achieves remarkable attacks against various MM algorithms. Our ablation study demonstrates that the proposed attack designs can progressively contribute to the attack performance. Finally, we show that prior defense mechanisms fail to defend against our attacks, highlighting the need for more advanced defense. Our code is available at: https://github.com/jzhang538/BadMerging.

Despite being prevalent in the general field of Natural Language Processing (NLP), pre-trained language models inherently carry privacy and copyright concerns due to their nature of training on large-scale web-scraped data. In this paper, we pioneer a systematic exploration of such risks associated with pre-trained language encoders, specifically focusing on the membership leakage of pre-training data exposed through downstream models adapted from pre-trained language encoders–an aspect largely overlooked in existing literature. Our study encompasses comprehensive experiments across four types of pre-trained encoder architectures, three representative downstream tasks, and five benchmark datasets. Intriguingly, our evaluations reveal, for the first time, the existence of membership leakage even when only the black-box output of the downstream model is exposed, highlighting a privacy risk far greater than previously assumed. Alongside, we present in-depth analysis and insights toward guiding future researchers and practitioners in addressing the privacy considerations in developing pre-trained language models.

Large vision-language models (LVLMs) have been rapidly developed and widely used in various fields, but the (potential) stereotypical bias in the model is largely unexplored.In this study, we present a pioneering measurement framework, ModSCAN, to SCAN the stereotypical bias within LVLMs from both vision and language Modalities.ModSCAN examines stereotypical biases with respect to two typical stereotypical attributes (gender and race) across three kinds of scenarios: occupations, descriptors, and persona traits.Our findings suggest that 1) the currently popular LVLMs show significant stereotype biases, with CogVLM emerging as the most biased model; 2) these stereotypical biases may stem from the inherent biases in the training dataset and pre-trained models; 3) the utilization of specific prompt prefixes (from both vision and language modalities) performs well in reducing stereotypical biases.We believe our work can serve as the foundation for understanding and addressing stereotypical bias in LVLMs.

Generative Adversarial Networks (GANs) have become widely used in model training, as they can improve performance and/or protect sensitive information by generating data. However, this also raises potential risks, as malicious GANs may compromise or sabotage models by poisoning their training data. Therefore, it is important to verify the origin of a model’s training data for accountability purposes. In this work, we take the first step in the forensic analysis of models trained on GAN-generated data. Specifically, we first detect whether a model is trained on GAN-generated or real data. We then attribute these models, trained on GAN-generated data, to their respective source GANs. We conduct extensive experiments on three datasets, using four popular GAN architectures and four common model architectures. Empirical results show the remarkable performance of our detection and attribution methods. Furthermore, we conduct a more in-depth study and reveal that models trained on various data sources exhibit different decision boundaries and behaviours.

Deep learning models, pivotal in AI applications, are susceptible to model hijacking attacks. In model hijacking attacks, adversaries can misuse models for unintended tasks, shifting blame and maintenance costs onto the models’ deployers. Existing attack methods re-purpose target models by poisoning their training sets during training. However, leading models like GPT-4 and BERT with vast parameters are often pruned before deployment on resource-limited devices, which presents challenges for in-training attacks, including existing model hi- jacking attacks. In this paper, we propose PRJack, the first pruning-resistant hijacking attack. Specifically, the adversary re-purposes a model to perform a hijacking task different from the original task, which can still be activated even after model pruning. Our experiments across multiple datasets and pruning techniques highlight PRJack’s remarkable superiority on pruned models over existing model hijacking attacks.

With the performance of deep neural networks (DNNs) remarkably improving, DNNs have been widely used in many areas. Consequently, the DNN model has become a valuable asset, and its intellectual property is safeguarded by ownership verification techniques (e.g., DNN fingerprinting). However, the feasibility of the DNN fingerprint removal attack and its potential influence remains an open problem. In this paper, we perform the first comprehensive investigation of DNN fingerprint removal attacks. Generally, the knowledge contained in a DNN model can be categorized into general semantic and fingerprint-specific knowledge. To this end, we propose a min-max bilevel optimization-based DNN fingerprint removal attack namedRemovalNet, to evade model ownership verification. The lower-level optimization is designed to remove fingerprint-specific knowledge. While in the upper-level optimization, we distill the victim model's general semantic knowledge to maintain the surrogate model's performance. We conduct extensive experiments to evaluate thefidelity,effectiveness, andefficiencyof theRemovalNetagainst four advanced defense methods on six metrics. The empirical results demonstrate that (1) theRemovalNetiseffective. After our DNN fingerprint removal attack, the model distance between the target and surrogate models is ×100 times higher than that of the baseline attacks, (2) theRemovalNetisefficient. It uses only 0.2% (400 samples) of the substitute dataset and 1,000 iterations to conduct our attack. Besides, compared with advanced model stealing attacks, theRemovalNetsaves nearly 85% of computational resources at most, (3) theRemovalNetachieves highfidelitythat the created surrogate model maintains high accuracy after the DNN fingerprint removal process.

Prompt-based learning is vulnerable to backdoor attacks.Existing backdoor attacks against prompt-based models consider injecting backdoors into the entire embedding layers or word embedding vectors.Such attacks can be easily affected by retraining on downstream tasks and with different prompting strategies, limiting the transferability of backdoor attacks.In this work, we propose transferable backdoor attacks against prompt-based models, called NOTABLE, which is independent of downstream tasks and prompting strategies.Specifically, NOTABLE injects backdoors into the encoders of PLMs by utilizing an adaptive verbalizer to bind triggers to specific words (i.e., anchors).It activates the backdoor by pasting input with triggers to reach adversary-desired anchors, achieving independence from downstream tasks and prompting strategies.We conduct experiments on six NLP tasks, three popular models, and three prompting strategies.Empirical results show that NOTABLE achieves superior attack performance (i.e., attack success rate over 90% on all the datasets), and outperforms two state-ofthe-art baselines.Evaluations on three defenses show the robustness of NOTABLE.Our code can be found at https://github.com/RU-System- Software-and-Security/Notable.

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·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.

Recently, the newly emerged multimodal models, which leverage both visual and linguistic modalities to train powerful encoders, have gained increasing attention. However, learning from a large-scale unlabeled dataset also exposes the model to the risk of potential poisoning attacks, whereby the adversary aims to perturb the model’s training data to trigger malicious behaviors in it. In contrast to previous work, only poisoning visual modality, in this work, we take the first step to studying poisoning attacks against multimodal models in both visual and linguistic modalities. Specially, we focus on answering two questions: (1) Is the linguistic modality also vulnerable to poisoning attacks? and (2) Which modality is most vulnerable? To answer the two questions, we propose three types of poisoning attacks against multimodal models. Extensive evaluations on different datasets and model architectures show that all three attacks can achieve significant attack performance while maintaining model utility in both visual and linguistic modalities. Furthermore, we observe that the poisoning effect differs between different modalities. To mitigate the attacks, we propose both pre-training and post-training defenses. We empirically show that both defenses can significantly reduce the attack performance while preserving the model’s utility. Our code is available at https://github.com/zqypku/mm_poison/.

Relying on the truth that not all inputs require the same level of computational cost to produce reliable predictions, multi-exit networks are gaining attention as a prominent approach for pushing the limits of efficient deployment. Multi-exit networks endow a backbone model with early exits, allowing predictions at intermediate layers of the model and thus saving computation time and energy. However, various current designs of multi-exit networks are only considered to achieve the best trade-off between resource usage efficiency and prediction accuracy, the privacy risks stemming from them have never been explored. This prompts the need for a comprehensive investigation of privacy risks in multi-exit networks.

Machine learning (ML) has been widely adopted in various privacy-critical applications, e.g., face recognition and medical image analysis. However, recent research has shown that ML models are vulnerable to attacks against their training data. Membership inference is one major attack in this domain: Given a data sample and model, an adversary aims to determine whether the sample is part of the model's training set. Existing membership inference attacks leverage the confidence scores returned by the model as their inputs (score-based attacks). However, these attacks can be easily mitigated if the model only exposes the predicted label, i.e., the final model decision. In this paper, we propose decision-based membership inference attacks and demonstrate that label-only exposures are also vulnerable to membership leakage. In particular, we develop two types of decision-based attacks, namely transfer attack and boundary attack. Empirical evaluation shows that our decision-based attacks can achieve remarkable performance, and even outperform the previous score-based attacks in some cases. We further present new insights on the success of membership inference based on quantitative and qualitative analysis, i.e., member samples of a model are more distant to the model's decision boundary than non-member samples. Finally, we evaluate multiple defense mechanisms against our decision-based attacks and show that our two types of attacks can bypass most of these defenses.

Deep learning techniques have made tremendous progress in a variety of challenging tasks, such as image recognition and machine translation, during the past decade. Training deep neural networks is computationally expensive and requires both human and intellectual resources. Therefore, it is necessary to protect the intellectual property of the model and externally verify the ownership of the model. However, previous studies either fail to defend against the evasion attack or have not explicitly dealt with fraudulent claims of ownership by adversaries. Furthermore, they can not establish a clear association between the model and the creator's identity.