隐私攻击

Although Large Language Models (LLMs) have demonstrated significant capabilities in executing complex tasks in a zero-shot manner, they are susceptible to jailbreak attacks and can be manipulated to produce harmful outputs. Recently, a growing body of research has categorized jailbreak attacks into token-level and prompt-level attacks. However, previous work primarily overlooks the diverse key factors of jailbreak attacks, with most studies concentrating on LLM vulnerabilities and lacking exploration of defense-enhanced LLMs. To address these issues, we introduced JailTrickBench to evaluate the impact of various attack settings on LLM performance and provide a baseline for jailbreak attacks, encouraging the adoption of a standardized evaluation framework. Specifically, we evaluate the eight key factors of implementing jailbreak attacks on LLMs from both target-level and attack-level perspectives. We further conduct seven representative jailbreak attacks on six defense methods across two widely used datasets, encompassing approximately 354 experiments with about 55,000 GPU hours on A800-80G. Our experimental results highlight the need for standardized benchmarking to evaluate these attacks on defense-enhanced LLMs. Our code is available at https://github.com/usail-hkust/JailTrickBench.

Non-interactive cryptographic computing, Fully Homomorphic Encryption (FHE), provides a promising solution for private neural network training on encrypted data. One challenge of FHE-based private training is its large computational overhead, especially the multiple rounds of forward and backward execution on each encrypted data sample. Considering the existence of largely redundant data samples, pruning them will significantly speed up the training, as proven in plain non-FHE training. Executing the data pruning of encrypted data on the server side is not trivial since the knowledge calculation of data pruning needs complex and expensive executions on encrypted data. There is a lack of FHE-based data pruning protocol for efficient, private training. In this paper, we propose, \textit{HEPrune}, to construct a FHE data-pruning protocol and then design an FHE-friendly data-pruning algorithm under client-aided or non-client-aided settings, respectively. We also observed that data sample pruning may not always remove ciphertexts, leaving large empty slots and limiting the effects of data pruning. Thus, in HEPrune, we further propose ciphertext-wise pruning to reduce ciphertext computation numbers without hurting accuracy. Experimental results show that our work can achieve a $16\times$ speedup with only a $0.6\%$ accuracy drop over prior work. The code is publicly available at \href{https://github.com/UCF-Lou-Lab-PET/Private-Data-Prune}.

Recent research has shown that representation learning models may accidentally memorize their training data. For example, the déjà vu method shows that for certain representation learning models and training images, it is sometimes possible to correctly predict the foreground label given only the representation of he background – better than through dataset-level correlations. However, their measurement method requires training two models – one to estimate dataset-level correlations and the other to estimate memorization. This multiple model setup becomes infeasible for large open-source models. In this work, we propose alter native simple methods to estimate dataset-level correlations, and show that these can be used to approximate an off-the-shelf model’s memorization ability without any retraining. This enables, for the first time, the measurement of memorization in pre-trained open-source image representation and vision-language models. Our results show that different ways of measuring memorization yield very similar aggregate results. We also find that open-source models typically have lower aggregate memorization than similar models trained on a subset of the data. The code is available both for vision (https://github.com/facebookresearch/DejaVuOSS) and vision language (https://github.com/facebookresearch/VLMDejaVu) models.

Imperceptible adversarial attacks aim to fool DNNs by adding imperceptible perturbation to the input data. Previous methods typically improve the imperceptibility of attacks by integrating common attack paradigms with specifically designed perception-based losses or the capabilities of generative models. In this paper, we propose Adversarial Attacks in Diffusion (AdvAD), a novel modeling framework distinct from existing attack paradigms. AdvAD innovatively conceptualizes attacking as a non-parametric diffusion process by theoretically exploring basic modeling approach rather than using the denoising or generation abilities of regular diffusion models requiring neural networks. At each step, much subtler yet effective adversarial guidance is crafted using only the attacked model without any additional network, which gradually leads the end of diffusion process from the original image to a desired imperceptible adversarial example. Grounded in a solid theoretical foundation of the proposed non-parametric diffusion process, AdvAD achieves high attack efficacy and imperceptibility with intrinsically lower overall perturbation strength. Additionally, an enhanced version AdvAD-X is proposed to evaluate the extreme of our novel framework under an ideal scenario. Extensive experiments demonstrate the effectiveness of the proposed AdvAD and AdvAD-X. Compared with state-of-the-art imperceptible attacks, AdvAD achieves an average of 99.9% (+17.3%) ASR with 1.34 (-0.97) $l_2$ distance, 49.74 (+4.76) PSNR and 0.9971 (+0.0043) SSIM against four prevalent DNNs with three different architectures on the ImageNet-compatible dataset. Code is available at https://github.com/XianguiKang/AdvAD.

We investigate the radioactivity of text generated by large language models (LLM), \ie whether it is possible to detect that such synthetic input was used to train a subsequent LLM. Current methods like membership inference or active IP protection either work only in settings where the suspected text is known or do not provide reliable statistical guarantees. We discover that, on the contrary, it is possible to reliably determine if a language model was trained on synthetic data if that data is output by a watermarked LLM. Our new methods, specialized for radioactivity, detects with a provable confidence weak residuals of the watermark signal in the fine-tuned LLM. We link the radioactivity contamination level to the following properties: the watermark robustness, its proportion in the training set, and the fine-tuning process. For instance, if the suspect model is open-weight, we demonstrate that training on watermarked instructions can be detected with high confidence ($p$-value $< 10^{-5}$) even when as little as $5\%$ of training text is watermarked.

We reveal the theoretical foundations of techniques for editing large language models, and present new methods which can do so without requiring retraining. Our theoretical insights show that a single metric (a measure of the intrinsic dimension of the model's features) can be used to assess a model's editability and reveals its previously unrecognised susceptibility to malicious *stealth attacks*. This metric is fundamental to predicting the success of a variety of editing approaches, and reveals new bridges between disparate families of editing methods. We collectively refer to these as *stealth editing* methods, because they directly update a model's weights to specify its response to specific known hallucinating prompts without affecting other model behaviour. By carefully applying our theoretical insights, we are able to introduce a new *jet-pack* network block which is optimised for highly selective model editing, uses only standard network operations, and can be inserted into existing networks. We also reveal the vulnerability of language models to stealth attacks: a small change to a model's weights which fixes its response to a single attacker-chosen prompt. Stealth attacks are computationally simple, do not require access to or knowledge of the model's training data, and therefore represent a potent yet previously unrecognised threat to redistributed foundation models. Extensive experimental results illustrate and support our methods and their theoretical underpinnings. Demos and source code are available at https://github.com/qinghua-zhou/stealth-edits.

In model inversion attacks (MIAs), adversaries attempt to recover private training data by exploiting access to a well-trained target model. Recent advancements have improved MIA performance using a two-stage generative framework. This approach first employs a generative adversarial network to learn a fixed distributional prior, which is then used to guide the inversion process during the attack. However, in this paper, we observed a phenomenon that such a fixed prior would lead to a low probability of sampling actual private data during the inversion process due to the inherent distribution gap between the prior distribution and the private data distribution, thereby constraining attack performance. To address this limitation, we propose increasing the density around high-quality pseudo-private data—recovered samples through model inversion that exhibit characteristics of the private training data—by slightly tuning the generator. This strategy effectively increases the probability of sampling actual private data that is close to these pseudo-private data during the inversion process. After integrating our method, the generative model inversion pipeline is strengthened, leading to improvements over state-of-the-art MIAs. This paves the way for new research directions in generative MIAs.

Deep neural networks have been known to be vulnerable to adversarial examples, which are inputs that are modified slightly to fool the network into making incorrect predictions. This has led to a significant amount of research on evaluating the robustness of these networks against such perturbations. One particularly important robustness metric is the robustness to minimal $\ell_{2}$ adversarial perturbations. However, existing methods for evaluating this robustness metric are either computationally expensive or not very accurate. In this paper, we introduce a new family of adversarial attacks that strike a balance between effectiveness and computational efficiency. Our proposed attacks are generalizations of the well-known DeepFool (DF) attack, while they remain simple to understand and implement. We demonstrate that our attacks outperform existing methods in terms of both effectiveness and computational efficiency. Our proposed attacks are also suitable for evaluating the robustness of large models and can be used to perform adversarial training (AT) to achieve state-of-the-art robustness to minimal $\ell_{2}$ adversarial perturbations.

Backdoor attacks aim to inject a backdoor into a classifier such that it predicts any input with an attacker-chosen backdoor trigger as an attacker-chosen target class. Existing backdoor attacks require either retraining the classifier with some clean data or modifying the model's architecture. As a result, they are 1) not applicable when clean data is unavailable, 2) less efficient when the model is large, and 3) less stealthy due to architecture changes. In this work, we propose DFBA, a novel retraining-free and data-free backdoor attack without changing the model architecture. Technically, our proposed method modifies a few parameters of a classifier to inject a backdoor. Through theoretical analysis, we verify that our injected backdoor is provably undetectable and unremovable by various state-of-the-art defenses under mild assumptions. Our evaluation on multiple datasets further demonstrates that our injected backdoor: 1) incurs negligible classification loss, 2) achieves 100\% attack success rates, and 3) bypasses six existing state-of-the-art defenses. Moreover, our comparison with a state-of-the-art non-data-free backdoor attack shows our attack is more stealthy and effective against various defenses while achieving less classification accuracy loss. We will release our code upon paper acceptance.

Safeguarding intellectual property and preventing potential misuse of AI-generated images are of paramount importance. This paper introduces a robust and agile plug-and-play watermark detection framework, referred to as RAW. As a departure from existing encoder-decoder methods, which incorporate fixed binary codes as watermarks within latent representations, our approach introduces learnable watermarks directly into the original image data. Subsequently, we employ a classifier that is jointly trained with the watermark to detect the presence of the watermark. The proposed framework is compatible with various generative architectures and supports on-the-fly watermark injection after training. By incorporating state-of-the-art smoothing techniques, we show that the framework also provides provable guarantees regarding the false positive rate for misclassifying a watermarked image, even in the presence of adversarial attacks targeting watermark removal. Experiments on a diverse range of images generated by state-of-the-art diffusion models demonstrate substantially improved watermark encoding speed and watermark detection performance, under adversarial attacks, while maintaining image quality. Our code is publicly available [here](https://github.com/jeremyxianx/RAWatermark).

Diffusion models have revolutionized customized text-to-image generation, allowing for efficient synthesis of photos from personal data with textual descriptions. However, these advancements bring forth risks including privacy breaches and unauthorized replication of artworks. Previous researches primarily center around using “prompt-specific methods” to generate adversarial examples to protect personal images, yet the effectiveness of existing methods is hindered by constrained adaptability to different prompts. In this paper, we introduce a Prompt-Agnostic Adversarial Perturbation (PAP) method for customized diffusion models. PAP first models the prompt distribution using a Laplace Approximation, and then produces prompt-agnostic perturbations by maximizing a disturbance expectation based on the modeled distribution. This approach effectively tackles the prompt-agnostic attacks, leading to improved defense stability. Extensive experiments in face privacy and artistic style protection, demonstrate the superior generalization of our method in comparison to existing techniques.

Large vision-language models (VLLMs) exhibit promising capabilities for processing multi-modal tasks across various application scenarios. However, their emergence also raises significant data security concerns, given the potential inclusion of sensitive information, such as private photos and medical records, in their training datasets. Detecting inappropriately used data in VLLMs remains a critical and unresolved issue, mainly due to the lack of standardized datasets and suitable methodologies. In this study, we introduce the first membership inference attack (MIA) benchmark tailored for various VLLMs to facilitate training data detection. Then, we propose a novel MIA pipeline specifically designed for token-level image detection. Lastly, we present a new metric called MaxRényi-K%, which is based on the confidence of the model output and applies to both text and image data. We believe that our work can deepen the understanding and methodology of MIAs in the context of VLLMs. Our code and datasets are available at https://github.com/LIONS-EPFL/VL-MIA.

We introduce WildTeaming, an automatic red-teaming framework that mines in-the-wild user-chatbot interactions to discover 5.7K unique clusters of novel jailbreak tactics, and then composes selections of multiple mined tactics for systematic exploration of novel and even more challenging jailbreaks. Compared to prior work that performed red-teaming via recruited human workers, gradient-based optimization, or iterative revision with large language models (LLMs), our work investigates jailbreaks from chatbot users in-the-wild who were not specifically instructed to break the system. WildTeaming reveals previously unidentified vulnerabilities of frontier LLMs, resulting in more diverse and successful adversarial attacks compared to state-of-the-art jailbreaking methods. While there exist many datasets for jailbreak evaluation, very few open-source datasets exist for jailbreak training, as safety training data has been closed among all frontier models even when their weights are open. Therefore, with WildTeaming we create WildJailbreak, a large-scale open-source synthetic safety dataset with 262K vanilla (direct request) and adversarial (complex jailbreak) prompt-response pairs. In order to mitigate exaggerated safety behaviors, WildJailbreak provides two contrastive types of queries: 1) harmful queries (both vanilla and adversarial) and 2) benign queries that resemble harmful queries in form but contain no harmful intent. As WildJailbreak considerably upgrades the quality and scale of existing safety resources, it uniquely enables us to examine the scaling effects of data and the interplay of data properties and model capabilities during safety training. Through extensive model training and evaluations, we identify the training properties that enable an ideal balance of safety behaviors: appropriate safeguarding without over-refusal, effective handling of both vanilla and adversarial queries, and minimal, if any, decrease in general capabilities. All the components of WildJailbreak contribute to achieving balanced safety behaviors of models

Scanning for trojan (backdoor) in deep neural networks is crucial due to their significant real-world applications. There has been an increasing focus on developing effective general trojan scanning methods across various trojan attacks. Despite advancements, there remains a shortage of methods that perform effectively without preconceived assumptions about the backdoor attack method. Additionally, we have observed that current methods struggle to identify classifiers trojaned using adversarial training. Motivated by these challenges, our study introduces a novel scanning method named TRODO (TROjan scanning by Detection of adversarial shifts in Out-of-distribution samples). TRODO leverages the concept of "blind spots"—regions where trojaned classifiers erroneously identify out-of-distribution (OOD) samples as in-distribution (ID). We scan for these blind spots by adversarially shifting OOD samples towards in-distribution. The increased likelihood of perturbed OOD samples being classified as ID serves as a signature for trojan detection. TRODO is both trojan and label mapping agnostic, effective even against adversarially trained trojaned classifiers. It is applicable even in scenarios where training data is absent, demonstrating high accuracy and adaptability across various scenarios and datasets, highlighting its potential as a robust trojan scanning strategy.

AI agents aim to solve complex tasks by combining text-based reasoning with external tool calls. Unfortunately, AI agents are vulnerable to prompt injection attacks where data returned by external tools hijacks the agent to execute malicious tasks. To measure the adversarial robustness of AI agents, we introduce AgentDojo, an evaluation framework for agents that execute tools over untrusted data. To capture the evolving nature of attacks and defenses, AgentDojo is not a static test suite, but rather an extensible environment for designing and evaluating new agent tasks, defenses, and adaptive attacks. We populate the environment with 97 realistic tasks (e.g., managing an email client, navigating an e-banking website, or making travel bookings), 629 security test cases, and various attack and defense paradigms from the literature. We find that AgentDojo poses a challenge for both attacks and defenses: state-of-the-art LLMs fail at many tasks (even in the absence of attacks), and existing prompt injection attacks break some security properties but not all. We hope that AgentDojo can foster research on new design principles for AI agents that solve common tasks in a reliable and robust manner.

Federated learning is a framework for collaborative machine learning where clients only share gradient updates and not their private data with a server. However, it was recently shown that gradient inversion attacks can reconstruct this data from the shared gradients. In the important honest-but-curious setting, existing attacks enable exact reconstruction only for batch size of $b=1$, with larger batches permitting only approximate reconstruction. In this work, we propose SPEAR, *the first algorithm reconstructing whole batches with $b >1$ exactly*. SPEAR combines insights into the explicit low-rank structure of gradients with a sampling-based algorithm. Crucially, we leverage ReLU-induced gradient sparsity to precisely filter out large numbers of incorrect samples, making a final reconstruction step tractable. We provide an efficient GPU implementation for fully connected networks and show that it recovers high-dimensional ImageNet inputs in batches of up to $b \lesssim 25$ exactly while scaling to large networks. Finally, we show theoretically that much larger batches can be reconstructed with high probability given exponential time.

When releasing outputs from confidential data, agencies need to balance the analytical usefulness of the released data with the obligation to protect data subjects' confidentiality. For releases satisfying differential privacy, this balance is reflected by the privacy budget, $\varepsilon$. We provide a framework for setting $\varepsilon$ based on its relationship with Bayesian posterior probabilities of disclosure. The agency responsible for the data release decides how much posterior risk it is willing to accept at various levels of prior risk, which implies a unique $\varepsilon$. Agencies can evaluate different risk profiles to determine one that leads to an acceptable trade-off in risk and utility.

Federated learning (FL) allows multiple devices to train a model collaboratively without sharing their data. Despite its benefits, FL is vulnerable to privacy leakage and poisoning attacks. To address the privacy concern, secure aggregation (SecAgg) is often used to obtain the aggregation of gradients on sever without inspecting individual user updates. Unfortunately, existing defense strategies against poisoning attacks rely on the analysis of local updates in plaintext, making them incompatible with SecAgg. To reconcile the conflicts, we propose a robust federated learning framework against poisoning attacks (RFLPA) based on SecAgg protocol. Our framework computes the cosine similarity between local updates and server updates to conduct robust aggregation. Furthermore, we leverage verifiable packed Shamir secret sharing to achieve reduced communication cost of $O(M+N)$ per user, and design a novel dot product aggregation algorithm to resolve the issue of increased information leakage. Our experimental results show that RFLPA significantly reduces communication and computation overhead by over $75\%$ compared to the state-of-the-art secret sharing method, BREA, while maintaining competitive accuracy.

Despite advances in AI alignment, large language models (LLMs) remain vulnerable to adversarial attacks or jailbreaking, in which adversaries can modify prompts to induce unwanted behavior. While some defenses have been proposed, they have not been adapted to newly proposed attacks and more challenging threat models. To address this, we propose an optimization-based objective for defending LLMs against jailbreaking attacks and an algorithm, Robust Prompt Optimization (RPO), to create robust system-level defenses. Our approach directly incorporates the adversary into the defensive objective and optimizes a lightweight and transferable suffix, enabling RPO to adapt to worst-case adaptive attacks. Our theoretical and experimental results show improved robustness to both jailbreaks seen during optimization and unknown jailbreaks, reducing the attack success rate (ASR) on GPT-4 to 6% and Llama-2 to 0% on JailbreakBench, setting the state-of-the-art.

Machine unlearning is motivated by principles of data autonomy. The premise is that a person can request to have their data's influence removed from deployed models, and those models should be updated as if they were retrained without the person's data. We show that these updates expose individuals to high-accuracy reconstruction attacks which allow the attacker to recover their data in its entirety, even when the original models are so simple that privacy risk might not otherwise have been a concern. We show how to mount a near-perfect attack on the deleted data point from linear regression models. We then generalize our attack to other loss functions and architectures, and empirically demonstrate the effectiveness of our attacks across a wide range of datasets (capturing both tabular and image data). Our work highlights that privacy risk is significant even for extremely simple model classes when individuals can request deletion of their data from the model.