Xinlei He

Xinlei He, Xinyue Shen, Zeyuan Chen et al.

Nowadays, powerful large language models (LLMs) such as ChatGPT have demonstrated revolutionary power in a variety of tasks. Consequently, the detection of machine-generated texts (MGTs) is becoming increasingly crucial as LLMs become more advanced and prevalent. These models have the ability to generate human-like language, making it challenging to discern whether a text is authored by a human or a machine. This raises concerns regarding authenticity, accountability, and potential bias. However, existing methods for detecting MGTs are evaluated using different model architectures, datasets, and experimental settings, resulting in a lack of a comprehensive evaluation framework that encompasses various methodologies. Furthermore, it remains unclear how existing detection methods would perform against powerful LLMs. In this paper, we fill this gap by proposing the first benchmark framework for MGT detection against powerful LLMs, named MGTBench. Extensive evaluations on public datasets with curated texts generated by various powerful LLMs such as ChatGPT-turbo and Claude demonstrate the effectiveness of different detection methods. Our ablation study shows that a larger number of words in general leads to better performance and most detection methods can achieve similar performance with much fewer training samples. Moreover, we delve into a more challenging task: text attribution. Our findings indicate that the model-based detection methods still perform well in the text attribution task. To investigate the robustness of different detection methods, we consider three adversarial attacks, namely paraphrasing, random spacing, and adversarial perturbations. We discover that these attacks can significantly diminish detection effectiveness, underscoring the critical need for the development of more robust detection methods.

Ziqing Yang, Xinlei He, Zheng Li et al.

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.

Zheng Li, Yiyong Liu, Xinlei He et al.

Relying on the fact that not all inputs require the same amount of computation to yield a confident prediction, 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 to obtain predictions at intermediate layers of the model and thus save computation time and/or energy. However, current various 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. In this paper, we perform the first privacy analysis of multi-exit networks through the lens of membership leakages. In particular, we first leverage the existing attack methodologies to quantify the multi-exit networks' vulnerability to membership leakages. Our experimental results show that multi-exit networks are less vulnerable to membership leakages and the exit (number and depth) attached to the backbone model is highly correlated with the attack performance. Furthermore, we propose a hybrid attack that exploits the exit information to improve the performance of existing attacks. We evaluate membership leakage threat caused by our hybrid attack under three different adversarial setups, ultimately arriving at a model-free and data-free adversary. These results clearly demonstrate that our hybrid attacks are very broadly applicable, thereby the corresponding risks are much more severe than shown by existing membership inference attacks. We further present a defense mechanism called TimeGuard specifically for multi-exit networks and show that TimeGuard mitigates the newly proposed attacks perfectly.

Yiting Qu, Xinlei He, Shannon Pierson et al.

The dissemination of hateful memes online has adverse effects on social media platforms and the real world. Detecting hateful memes is challenging, one of the reasons being the evolutionary nature of memes; new hateful memes can emerge by fusing hateful connotations with other cultural ideas or symbols. In this paper, we propose a framework that leverages multimodal contrastive learning models, in particular OpenAI's CLIP, to identify targets of hateful content and systematically investigate the evolution of hateful memes. We find that semantic regularities exist in CLIP-generated embeddings that describe semantic relationships within the same modality (images) or across modalities (images and text). Leveraging this property, we study how hateful memes are created by combining visual elements from multiple images or fusing textual information with a hateful image. We demonstrate the capabilities of our framework for analyzing the evolution of hateful memes by focusing on antisemitic memes, particularly the Happy Merchant meme. Using our framework on a dataset extracted from 4chan, we find 3.3K variants of the Happy Merchant meme, with some linked to specific countries, persons, or organizations. We envision that our framework can be used to aid human moderators by flagging new variants of hateful memes so that moderators can manually verify them and mitigate the problem of hateful content online.

Yihan Ma, Zhengyu Zhao, Xinlei He et al.

Large text-to-image models have shown remarkable performance in synthesizing high-quality images. In particular, the subject-driven model makes it possible to personalize the image synthesis for a specific subject, e.g., a human face or an artistic style, by fine-tuning the generic text-to-image model with a few images from that subject. Nevertheless, misuse of subject-driven image synthesis may violate the authority of subject owners. For example, malicious users may use subject-driven synthesis to mimic specific artistic styles or to create fake facial images without authorization. To protect subject owners against such misuse, recent attempts have commonly relied on adversarial examples to indiscriminately disrupt subject-driven image synthesis. However, this essentially prevents any benign use of subject-driven synthesis based on protected images. In this paper, we take a different angle and aim at protection without sacrificing the utility of protected images for general synthesis purposes. Specifically, we propose GenWatermark, a novel watermark system based on jointly learning a watermark generator and a detector. In particular, to help the watermark survive the subject-driven synthesis, we incorporate the synthesis process in learning GenWatermark by fine-tuning the detector with synthesized images for a specific subject. This operation is shown to largely improve the watermark detection accuracy and also ensure the uniqueness of the watermark for each individual subject. Extensive experiments validate the effectiveness of GenWatermark, especially in practical scenarios with unknown models and text prompts (74% Acc.), as well as partial data watermarking (80% Acc. for 1/4 watermarking). We also demonstrate the robustness of GenWatermark to two potential countermeasures that substantially degrade the synthesis quality.

Xinlei He, Savvas Zannettou, Yun Shen et al.

The spread of toxic content online is an important problem that has adverse effects on user experience online and in our society at large. Motivated by the importance and impact of the problem, research focuses on developing solutions to detect toxic content, usually leveraging machine learning (ML) models trained on human-annotated datasets. While these efforts are important, these models usually do not generalize well and they can not cope with new trends (e.g., the emergence of new toxic terms). Currently, we are witnessing a shift in the approach to tackling societal issues online, particularly leveraging large language models (LLMs) like GPT-3 or T5 that are trained on vast corpora and have strong generalizability. In this work, we investigate how we can use LLMs and prompt learning to tackle the problem of toxic content, particularly focusing on three tasks; 1) Toxicity Classification, 2) Toxic Span Detection, and 3) Detoxification. We perform an extensive evaluation over five model architectures and eight datasets demonstrating that LLMs with prompt learning can achieve similar or even better performance compared to models trained on these specific tasks. We find that prompt learning achieves around 10\% improvement in the toxicity classification task compared to the baselines, while for the toxic span detection task we find better performance to the best baseline (0.643 vs. 0.640 in terms of $F_1$-score). Finally, for the detoxification task, we find that prompt learning can successfully reduce the average toxicity score (from 0.775 to 0.213) while preserving semantic meaning.

Xinyue Shen, Xinlei He, Zheng Li et al.

Masked image modeling (MIM) revolutionizes self-supervised learning (SSL) for image pre-training. In contrast to previous dominating self-supervised methods, i.e., contrastive learning, MIM attains state-of-the-art performance by masking and reconstructing random patches of the input image. However, the associated security and privacy risks of this novel generative method are unexplored. In this paper, we perform the first security risk quantification of MIM through the lens of backdoor attacks. Different from previous work, we are the first to systematically threat modeling on SSL in every phase of the model supply chain, i.e., pre-training, release, and downstream phases. Our evaluation shows that models built with MIM are vulnerable to existing backdoor attacks in release and downstream phases and are compromised by our proposed method in pre-training phase. For instance, on CIFAR10, the attack success rate can reach 99.62%, 96.48%, and 98.89% in the downstream phase, release phase, and pre-training phase, respectively. We also take the first step to investigate the success factors of backdoor attacks in the pre-training phase and find the trigger number and trigger pattern play key roles in the success of backdoor attacks while trigger location has only tiny effects. In the end, our empirical study of the defense mechanisms across three detection-level on model supply chain phases indicates that different defenses are suitable for backdoor attacks in different phases. However, backdoor attacks in the release phase cannot be detected by all three detection-level methods, calling for more effective defenses in future research.

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.

Xinlei He, Hongbin Liu, Neil Zhenqiang Gong et al.

Semi-supervised learning (SSL) leverages both labeled and unlabeled data to train machine learning (ML) models. State-of-the-art SSL methods can achieve comparable performance to supervised learning by leveraging much fewer labeled data. However, most existing works focus on improving the performance of SSL. In this work, we take a different angle by studying the training data privacy of SSL. Specifically, we propose the first data augmentation-based membership inference attacks against ML models trained by SSL. Given a data sample and the black-box access to a model, the goal of membership inference attack is to determine whether the data sample belongs to the training dataset of the model. Our evaluation shows that the proposed attack can consistently outperform existing membership inference attacks and achieves the best performance against the model trained by SSL. Moreover, we uncover that the reason for membership leakage in SSL is different from the commonly believed one in supervised learning, i.e., overfitting (the gap between training and testing accuracy). We observe that the SSL model is well generalized to the testing data (with almost 0 overfitting) but ''memorizes'' the training data by giving a more confident prediction regardless of its correctness. We also explore early stopping as a countermeasure to prevent membership inference attacks against SSL. The results show that early stopping can mitigate the membership inference attack, but with the cost of model's utility degradation.

Boyang Zhang, Xinlei He, Yun Shen et al.

Building advanced machine learning (ML) models requires expert knowledge and many trials to discover the best architecture and hyperparameter settings. Previous work demonstrates that model information can be leveraged to assist other attacks, such as membership inference, generating adversarial examples. Therefore, such information, e.g., hyperparameters, should be kept confidential. It is well known that an adversary can leverage a target ML model's output to steal the model's information. In this paper, we discover a new side channel for model information stealing attacks, i.e., models' scientific plots which are extensively used to demonstrate model performance and are easily accessible. Our attack is simple and straightforward. We leverage the shadow model training techniques to generate training data for the attack model which is essentially an image classifier. Extensive evaluation on three benchmark datasets shows that our proposed attack can effectively infer the architecture/hyperparameters of image classifiers based on convolutional neural network (CNN) given the scientific plot generated from it. We also reveal that the attack's success is mainly caused by the shape of the scientific plots, and further demonstrate that the attacks are robust in various scenarios. Given the simplicity and effectiveness of the attack method, our study indicates scientific plots indeed constitute a valid side channel for model information stealing attacks. To mitigate the attacks, we propose several defense mechanisms that can reduce the original attacks' accuracy while maintaining the plot utility. However, such defenses can still be bypassed by adaptive attacks.

Xinlei He, Zheng Li, Weilin Xu et al.

Machine learning models are prone to memorizing sensitive data, making them vulnerable to membership inference attacks in which an adversary aims to infer whether an input sample was used to train the model. Over the past few years, researchers have produced many membership inference attacks and defenses. However, these attacks and defenses employ a variety of strategies and are conducted in different models and datasets. The lack of comprehensive benchmark, however, means we do not understand the strengths and weaknesses of existing attacks and defenses. We fill this gap by presenting a large-scale measurement of different membership inference attacks and defenses. We systematize membership inference through the study of nine attacks and six defenses and measure the performance of different attacks and defenses in the holistic evaluation. We then quantify the impact of the threat model on the results of these attacks. We find that some assumptions of the threat model, such as same-architecture and same-distribution between shadow and target models, are unnecessary. We are also the first to execute attacks on the real-world data collected from the Internet, instead of laboratory datasets. We further investigate what determines the performance of membership inference attacks and reveal that the commonly believed overfitting level is not sufficient for the success of the attacks. Instead, the Jensen-Shannon distance of entropy/cross-entropy between member and non-member samples correlates with attack performance much better. This gives us a new way to accurately predict membership inference risks without running the attack. Finally, we find that data augmentation degrades the performance of existing attacks to a larger extent, and we propose an adaptive attack using augmentation to train shadow and attack models that improve attack performance.

Yihan Ma, Zhikun Zhang, Ning Yu et al.

Graph generative models become increasingly effective for data distribution approximation and data augmentation. While they have aroused public concerns about their malicious misuses or misinformation broadcasts, just as what Deepfake visual and auditory media has been delivering to society. Hence it is essential to regulate the prevalence of generated graphs. To tackle this problem, we pioneer the formulation of the generated graph detection problem to distinguish generated graphs from real ones. We propose the first framework to systematically investigate a set of sophisticated models and their performance in four classification scenarios. Each scenario switches between seen and unseen datasets/generators during testing to get closer to real-world settings and progressively challenge the classifiers. Extensive experiments evidence that all the models are qualified for generated graph detection, with specific models having advantages in specific scenarios. Resulting from the validated generality and oblivion of the classifiers to unseen datasets/generators, we draw a safe conclusion that our solution can sustain for a decent while to curb generated graph misuses.

Boyang Zhang, Zheng Li, Ziqing Yang et al.

While advanced machine learning (ML) models are deployed in numerous real-world applications, previous works demonstrate these models have security and privacy vulnerabilities. Various empirical research has been done in this field. However, most of the experiments are performed on target ML models trained by the security researchers themselves. Due to the high computational resource requirement for training advanced models with complex architectures, researchers generally choose to train a few target models using relatively simple architectures on typical experiment datasets. We argue that to understand ML models' vulnerabilities comprehensively, experiments should be performed on a large set of models trained with various purposes (not just the purpose of evaluating ML attacks and defenses). To this end, we propose using publicly available models with weights from the Internet (public models) for evaluating attacks and defenses on ML models. We establish a database, namely SecurityNet, containing 910 annotated image classification models. We then analyze the effectiveness of several representative attacks/defenses, including model stealing attacks, membership inference attacks, and backdoor detection on these public models. Our evaluation empirically shows the performance of these attacks/defenses can vary significantly on public models compared to self-trained models. We share SecurityNet with the research community. and advocate researchers to perform experiments on public models to better demonstrate their proposed methods' effectiveness in the future.

Zheng Li, Xinlei He, Ning Yu et al.

Masked Image Modeling (MIM) has achieved significant success in the realm of self-supervised learning (SSL) for visual recognition. The image encoder pre-trained through MIM, involving the masking and subsequent reconstruction of input images, attains state-of-the-art performance in various downstream vision tasks. However, most existing works focus on improving the performance of MIM.In this work, we take a different angle by studying the pre-training data privacy of MIM. Specifically, we propose the first membership inference attack against image encoders pre-trained by MIM, which aims to determine whether an image is part of the MIM pre-training dataset. The key design is to simulate the pre-training paradigm of MIM, i.e., image masking and subsequent reconstruction, and then obtain reconstruction errors. These reconstruction errors can serve as membership signals for achieving attack goals, as the encoder is more capable of reconstructing the input image in its training set with lower errors. Extensive evaluations are conducted on three model architectures and three benchmark datasets. Empirical results show that our attack outperforms baseline methods. Additionally, we undertake intricate ablation studies to analyze multiple factors that could influence the performance of the attack.

Zhen Sun, Zongmin Zhang, Xinyue Shen et al.

Social media platforms are experiencing a growing presence of AI-Generated Texts (AIGTs). However, the misuse of AIGTs could have profound implications for public opinion, such as spreading misinformation and manipulating narratives. Despite its importance, it remains unclear how prevalent AIGTs are on social media. To address this gap, this paper aims to quantify and monitor the AIGTs on online social media platforms. We first collect a dataset (SM-D) with around 2.4M posts from 3 major social media platforms: Medium, Quora, and Reddit. Then, we construct a diverse dataset (AIGTBench) to train and evaluate AIGT detectors. AIGTBench combines popular open-source datasets and our AIGT datasets generated from social media texts by 12 LLMs, serving as a benchmark for evaluating mainstream detectors. With this setup, we identify the best-performing detector (OSM-Det). We then apply OSM-Det to SM-D to track AIGTs across social media platforms from January 2022 to October 2024, using the AI Attribution Rate (AAR) as the metric. Specifically, Medium and Quora exhibit marked increases in AAR, rising from 1.77% to 37.03% and 2.06% to 38.95%, respectively. In contrast, Reddit shows slower growth, with AAR increasing from 1.31% to 2.45% over the same period. Our further analysis indicates that AIGTs on social media differ from human-written texts across several dimensions, including linguistic patterns, topic distributions, engagement levels, and the follower distribution of authors. We envision our analysis and findings on AIGTs in social media can shed light on future research in this domain.

Jingyi Zheng, Tianyi Hu, Tianshuo Cong et al.

Backdoor attacks significantly compromise the security of large language models by triggering them to output specific and controlled content. Currently, triggers for textual backdoor attacks fall into two categories: fixed-token triggers and sentence-pattern triggers. However, the former are typically easy to identify and filter, while the latter, such as syntax and style, do not apply to all original samples and may lead to semantic shifts. In this paper, inspired by cross-lingual (CL) prompts of LLMs in real-world scenarios, we propose a higher-dimensional trigger method at the paragraph level, namely CL-attack. CL-attack injects the backdoor by using texts with specific structures that incorporate multiple languages, thereby offering greater stealthiness and universality compared to existing backdoor attack techniques. Extensive experiments on different tasks and model architectures demonstrate that CL-attack can achieve nearly 100% attack success rate with a low poisoning rate in both classification and generation tasks. We also empirically show that the CL-attack is more robust against current major defense methods compared to baseline backdoor attacks. Additionally, to mitigate CL-attack, we further develop a new defense called TranslateDefense, which can partially mitigate the impact of CL-attack.

Zeren Luo, Zifan Peng, Yule Liu et al.

Recent advancements in Large Language Models (LLMs) have significantly enhanced the capabilities of AI-Powered Search Engines (AIPSEs), offering precise and efficient responses by integrating external databases with pre-existing knowledge. However, we observe that these AIPSEs raise risks such as quoting malicious content or citing malicious websites, leading to harmful or unverified information dissemination. In this study, we conduct the first safety risk quantification on seven production AIPSEs by systematically defining the threat model, risk type, and evaluating responses to various query types. With data collected from PhishTank, ThreatBook, and LevelBlue, our findings reveal that AIPSEs frequently generate harmful content that contains malicious URLs even with benign queries (e.g., with benign keywords). We also observe that directly querying a URL will increase the number of main risk-inclusive responses, while querying with natural language will slightly mitigate such risk. Compared to traditional search engines, AIPSEs outperform in both utility and safety. We further perform two case studies on online document spoofing and phishing to show the ease of deceiving AIPSEs in the real-world setting. To mitigate these risks, we develop an agent-based defense with a GPT-4.1-based content refinement tool and a URL detector. Our evaluation shows that our defense can effectively reduce the risk, with only a minor cost of reducing available information by approximately 10.7%. Our research highlights the urgent need for robust safety measures in AIPSEs.

Jingyi Zheng, Zifan Peng, Yule Liu et al.

Smart contracts are trustworthy, immutable, and automatically executed programs on the blockchain. Their execution requires the Gas mechanism to ensure efficiency and fairness. However, due to non-optimal coding practices, many contracts contain Gas waste patterns that need to be optimized. Existing solutions mostly rely on manual discovery, which is inefficient, costly to maintain, and difficult to scale. Recent research uses large language models (LLMs) to explore new Gas waste patterns. However, it struggles to remain compatible with existing patterns, often produces redundant patterns, and requires manual validation/rewriting. To address this gap, we present GasAgent, the first multi-agent system for smart contract Gas optimization that combines compatibility with existing patterns and automated discovery/validation of new patterns, enabling end-to-end optimization. GasAgent consists of four specialized agents, Seeker, Innovator, Executor, and Manager, that collaborate in a closed loop to identify, validate, and apply Gas-saving improvements. Experiments on 100 verified real-world contracts demonstrate that GasAgent successfully optimizes 82 contracts, achieving an average deployment Gas savings of 9.97%. In addition, our evaluation confirms its compatibility with existing tools and validates the effectiveness of each module through ablation studies. To assess broader usability, we further evaluate 500 contracts generated by five representative LLMs across 10 categories and find that GasAgent optimizes 79.8% of them, with deployment Gas savings ranging from 4.79% to 13.93%, showing its usability as the optimization layer for LLM-assisted smart contract development.

Yiming Huang, Junyan Zhang, Zihao Wang et al.

Large Language Models (LLMs) have become powerful, but hallucinations remain a vital obstacle to their trustworthy use. While previous works improved the capability of hallucination detection by measuring uncertainty, they all lack the ability to explain the provenance behind why hallucinations occur, i.e., which part of the inputs tends to trigger hallucinations. Recent works on the prompt attack indicate that uncertainty exists in semantic propagation, where attention mechanisms gradually fuse local token information into high-level semantics across layers. Meanwhile, uncertainty also emerges in language generation, due to its probability-based selection of high-level semantics for sampled generations. Based on that, we propose RePPL to recalibrate uncertainty measurement by these two aspects, which dispatches explainable uncertainty scores to each token and aggregates in Perplexity-style Log-Average form as total score. Experiments show that our method achieves the best comprehensive detection performance across various QA datasets on advanced models (average AUC of 0.833), and our method is capable of producing token-level uncertainty scores as explanations for the hallucination. Leveraging these scores, we preliminarily find the chaotic pattern of hallucination and showcase its promising usage.

Delong Ran, Jinyuan Liu, Yichen Gong et al.

Jailbreak attacks induce Large Language Models (LLMs) to generate harmful responses, posing severe misuse threats. Though research on jailbreak attacks and defenses is emerging, there is no consensus on evaluating jailbreaks, i.e., the methods to assess the harmfulness of an LLM's response are varied. Each approach has its own set of strengths and weaknesses, impacting their alignment with human values, as well as the time and financial cost. This diversity challenges researchers in choosing suitable evaluation methods and comparing different attacks and defenses. In this paper, we conduct a comprehensive analysis of jailbreak evaluation methodologies, drawing from nearly 90 jailbreak research published between May 2023 and April 2024. Our study introduces a systematic taxonomy of jailbreak evaluators, offering indepth insights into their strengths and weaknesses, along with the current status of their adaptation. To aid further research, we propose JailbreakEval, a toolkit for evaluating jailbreak attempts. JailbreakEval includes various evaluators out-of-the-box, enabling users to obtain results with a single command or customized evaluation workflows. In summary, we regard JailbreakEval to be a catalyst that simplifies the evaluation process in jailbreak research and fosters an inclusive standard for jailbreak evaluation within the community.

Yanling Wang, Haoyang Li, Hao Zou et al.

Despite advancements in large language models (LLMs), non-factual responses still persist in fact-seeking question answering. Unlike extensive studies on post-hoc detection of these responses, this work studies non-factuality prediction (NFP), predicting whether an LLM will generate a non-factual response prior to the response generation. Previous NFP methods have shown LLMs' awareness of their knowledge, but they face challenges in terms of efficiency and transferability. In this work, we propose a lightweight model named Factuality Lens (FacLens), which effectively probes hidden representations of fact-seeking questions for the NFP task. Moreover, we discover that hidden question representations sourced from different LLMs exhibit similar NFP patterns, enabling the transferability of FacLens across different LLMs to reduce development costs. Extensive experiments highlight FacLens's superiority in both effectiveness and efficiency.

Yiting Qu, Xinyue Shen, Xinlei He et al.

State-of-the-art Text-to-Image models like Stable Diffusion and DALLE$\cdot$2 are revolutionizing how people generate visual content. At the same time, society has serious concerns about how adversaries can exploit such models to generate unsafe images. In this work, we focus on demystifying the generation of unsafe images and hateful memes from Text-to-Image models. We first construct a typology of unsafe images consisting of five categories (sexually explicit, violent, disturbing, hateful, and political). Then, we assess the proportion of unsafe images generated by four advanced Text-to-Image models using four prompt datasets. We find that these models can generate a substantial percentage of unsafe images; across four models and four prompt datasets, 14.56% of all generated images are unsafe. When comparing the four models, we find different risk levels, with Stable Diffusion being the most prone to generating unsafe content (18.92% of all generated images are unsafe). Given Stable Diffusion's tendency to generate more unsafe content, we evaluate its potential to generate hateful meme variants if exploited by an adversary to attack a specific individual or community. We employ three image editing methods, DreamBooth, Textual Inversion, and SDEdit, which are supported by Stable Diffusion. Our evaluation result shows that 24% of the generated images using DreamBooth are hateful meme variants that present the features of the original hateful meme and the target individual/community; these generated images are comparable to hateful meme variants collected from the real world. Overall, our results demonstrate that the danger of large-scale generation of unsafe images is imminent. We discuss several mitigating measures, such as curating training data, regulating prompts, and implementing safety filters, and encourage better safeguard tools to be developed to prevent unsafe generation.

Tianshuo Cong, Xinlei He, Yang Zhang

Self-supervised learning is an emerging machine learning paradigm. Compared to supervised learning which leverages high-quality labeled datasets, self-supervised learning relies on unlabeled datasets to pre-train powerful encoders which can then be treated as feature extractors for various downstream tasks. The huge amount of data and computational resources consumption makes the encoders themselves become the valuable intellectual property of the model owner. Recent research has shown that the machine learning model's copyright is threatened by model stealing attacks, which aim to train a surrogate model to mimic the behavior of a given model. We empirically show that pre-trained encoders are highly vulnerable to model stealing attacks. However, most of the current efforts of copyright protection algorithms such as watermarking concentrate on classifiers. Meanwhile, the intrinsic challenges of pre-trained encoder's copyright protection remain largely unstudied. We fill the gap by proposing SSLGuard, the first watermarking scheme for pre-trained encoders. Given a clean pre-trained encoder, SSLGuard injects a watermark into it and outputs a watermarked version. The shadow training technique is also applied to preserve the watermark under potential model stealing attacks. Our extensive evaluation shows that SSLGuard is effective in watermark injection and verification, and it is robust against model stealing and other watermark removal attacks such as input noising, output perturbing, overwriting, model pruning, and fine-tuning.

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.