Minzhou Pan

Minzhou Pan, Yi Zeng, Lingjuan Lyu et al.

Backdoor data detection is traditionally studied in an end-to-end supervised learning (SL) setting. However, recent years have seen the proliferating adoption of self-supervised learning (SSL) and transfer learning (TL), due to their lesser need for labeled data. Successful backdoor attacks have also been demonstrated in these new settings. However, we lack a thorough understanding of the applicability of existing detection methods across a variety of learning settings. By evaluating 56 attack settings, we show that the performance of most existing detection methods varies significantly across different attacks and poison ratios, and all fail on the state-of-the-art clean-label attack. In addition, they either become inapplicable or suffer large performance losses when applied to SSL and TL. We propose a new detection method called Active Separation via Offset (ASSET), which actively induces different model behaviors between the backdoor and clean samples to promote their separation. We also provide procedures to adaptively select the number of suspicious points to remove. In the end-to-end SL setting, ASSET is superior to existing methods in terms of consistency of defensive performance across different attacks and robustness to changes in poison ratios; in particular, it is the only method that can detect the state-of-the-art clean-label attack. Moreover, ASSET's average detection rates are higher than the best existing methods in SSL and TL, respectively, by 69.3% and 33.2%, thus providing the first practical backdoor defense for these new DL settings. We open-source the project to drive further development and encourage engagement: https://github.com/ruoxi-jia-group/ASSET.

Yi Zeng, Yu Yang, Andy Zhou et al. · stanford

Foundation models (FMs) provide societal benefits but also amplify risks. Governments, companies, and researchers have proposed regulatory frameworks, acceptable use policies, and safety benchmarks in response. However, existing public benchmarks often define safety categories based on previous literature, intuitions, or common sense, leading to disjointed sets of categories for risks specified in recent regulations and policies, which makes it challenging to evaluate and compare FMs across these benchmarks. To bridge this gap, we introduce AIR-Bench 2024, the first AI safety benchmark aligned with emerging government regulations and company policies, following the regulation-based safety categories grounded in our AI risks study, AIR 2024. AIR 2024 decomposes 8 government regulations and 16 company policies into a four-tiered safety taxonomy with 314 granular risk categories in the lowest tier. AIR-Bench 2024 contains 5,694 diverse prompts spanning these categories, with manual curation and human auditing to ensure quality. We evaluate leading language models on AIR-Bench 2024, uncovering insights into their alignment with specified safety concerns. By bridging the gap between public benchmarks and practical AI risks, AIR-Bench 2024 provides a foundation for assessing model safety across jurisdictions, fostering the development of safer and more responsible AI systems.

Yi Zeng, Minzhou Pan, Hoang Anh Just et al.

Backdoor attacks insert malicious data into a training set so that, during inference time, it misclassifies inputs that have been patched with a backdoor trigger as the malware specified label. For backdoor attacks to bypass human inspection, it is essential that the injected data appear to be correctly labeled. The attacks with such property are often referred to as "clean-label attacks." Existing clean-label backdoor attacks require knowledge of the entire training set to be effective. Obtaining such knowledge is difficult or impossible because training data are often gathered from multiple sources (e.g., face images from different users). It remains a question whether backdoor attacks still present a real threat. This paper provides an affirmative answer to this question by designing an algorithm to mount clean-label backdoor attacks based only on the knowledge of representative examples from the target class. With poisoning equal to or less than 0.5% of the target-class data and 0.05% of the training set, we can train a model to classify test examples from arbitrary classes into the target class when the examples are patched with a backdoor trigger. Our attack works well across datasets and models, even when the trigger presents in the physical world. We explore the space of defenses and find that, surprisingly, our attack can evade the latest state-of-the-art defenses in their vanilla form, or after a simple twist, we can adapt to the downstream defenses. We study the cause of the intriguing effectiveness and find that because the trigger synthesized by our attack contains features as persistent as the original semantic features of the target class, any attempt to remove such triggers would inevitably hurt the model accuracy first.

Yi Zeng, Minzhou Pan, Himanshu Jahagirdar et al.

Given the volume of data needed to train modern machine learning models, external suppliers are increasingly used. However, incorporating external data poses data poisoning risks, wherein attackers manipulate their data to degrade model utility or integrity. Most poisoning defenses presume access to a set of clean data (or base set). While this assumption has been taken for granted, given the fast-growing research on stealthy poisoning attacks, a question arises: can defenders really identify a clean subset within a contaminated dataset to support defenses? This paper starts by examining the impact of poisoned samples on defenses when they are mistakenly mixed into the base set. We analyze five defenses and find that their performance deteriorates dramatically with less than 1% poisoned points in the base set. These findings suggest that sifting out a base set with high precision is key to these defenses' performance. Motivated by these observations, we study how precise existing automated tools and human inspection are at identifying clean data in the presence of data poisoning. Unfortunately, neither effort achieves the precision needed. Worse yet, many of the outcomes are worse than random selection. In addition to uncovering the challenge, we propose a practical countermeasure, Meta-Sift. Our method is based on the insight that existing attacks' poisoned samples shifts from clean data distributions. Hence, training on the clean portion of a dataset and testing on the corrupted portion will result in high prediction loss. Leveraging the insight, we formulate a bilevel optimization to identify clean data and further introduce a suite of techniques to improve efficiency and precision. Our evaluation shows that Meta-Sift can sift a clean base set with 100% precision under a wide range of poisoning attacks. The selected base set is large enough to give rise to successful defenses.

Xu Li, Simon Yu, Minzhou Pan et al.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

Zhenting Wang, Chen Chen, Vikash Sehwag et al.

The popularity of visual generative AI models like DALL-E 3, Stable Diffusion XL, Stable Video Diffusion, and Sora has been increasing. Through extensive evaluation, we discovered that the state-of-the-art visual generative models can generate content that bears a striking resemblance to characters protected by intellectual property rights held by major entertainment companies (such as Sony, Marvel, and Nintendo), which raises potential legal concerns. This happens when the input prompt contains the character's name or even just descriptive details about their characteristics. To mitigate such IP infringement problems, we also propose a defense method against it. In detail, we develop a revised generation paradigm that can identify potentially infringing generated content and prevent IP infringement by utilizing guidance techniques during the diffusion process. It has the capability to recognize generated content that may be infringing on intellectual property rights, and mitigate such infringement by employing guidance methods throughout the diffusion process without retrain or fine-tune the pretrained models. Experiments on well-known character IPs like Spider-Man, Iron Man, and Superman demonstrate the effectiveness of the proposed defense method. Our data and code can be found at https://github.com/ZhentingWang/GAI_IP_Infringement.

Jingxu Ng, Cheng Lv, Pu Zhao et al.

Stable diffusion plays a crucial role in generating high-quality images. However, image generation is time-consuming and memory-intensive. To address this, stable-diffusion.cpp (Sdcpp) emerges as an efficient inference framework to accelerate the diffusion models. Although it is lightweight, the current implementation of ggml_conv_2d operator in Sdcpp is suboptimal, exhibiting both high inference latency and massive memory usage. To address this, in this work, we present an optimized version of Sdcpp leveraging the Winograd algorithm to accelerate 2D convolution operations, which is the primary bottleneck in the pipeline. By analyzing both dependent and independent computation graphs, we exploit the device's locality and parallelism to achieve substantial performance improvements. Our framework delivers correct end-to-end results across various stable diffusion models, including SDv1.4, v1.5, v2.1, SDXL, and SDXL-Turbo. Our evaluation results demonstrate a speedup up to 2.76x for individual convolutional layers and an inference speedup up to 4.79x for the overall image generation process, compared with the original Sdcpp on M1 pro. Homepage: https://github.com/SealAILab/stable-diffusion-cpp

Zhaorun Chen, Francesco Pinto, Minzhou Pan et al.

With the rise of generative AI and rapid growth of high-quality video generation, video guardrails have become more crucial than ever to ensure safety and security across platforms. Current video guardrails, however, are either overly simplistic, relying on pure classification models trained on simple policies with limited unsafe categories, which lack detailed explanations, or prompting multimodal large language models (MLLMs) with long safety guidelines, which are inefficient and impractical for guardrailing real-world content. To bridge this gap, we propose SafeWatch, an efficient MLLM-based video guardrail model designed to follow customized safety policies and provide multi-label video guardrail outputs with content-specific explanations in a zero-shot manner. In particular, unlike traditional MLLM-based guardrails that encode all safety policies autoregressively, causing inefficiency and bias, SafeWatch uniquely encodes each policy chunk in parallel and eliminates their position bias such that all policies are attended simultaneously with equal importance. In addition, to improve efficiency and accuracy, SafeWatch incorporates a policy-aware visual token pruning algorithm that adaptively selects the most relevant video tokens for each policy, discarding noisy or irrelevant information. This allows for more focused, policy-compliant guardrail with significantly reduced computational overhead. Considering the limitations of existing video guardrail benchmarks, we propose SafeWatch-Bench, a large-scale video guardrail benchmark comprising over 2M videos spanning six safety categories which covers over 30 tasks to ensure a comprehensive coverage of all potential safety scenarios. SafeWatch outperforms SOTA by 28.2% on SafeWatch-Bench, 13.6% on benchmarks, cuts costs by 10%, and delivers top-tier explanations validated by LLM and human reviews.

Minzhou Pan, Zhenting Wang, Xin Dong et al. · princeton

In this paper, we propose WaterMark Detection (WMD), the first invisible watermark detection method under a black-box and annotation-free setting. WMD is capable of detecting arbitrary watermarks within a given reference dataset using a clean non-watermarked dataset as a reference, without relying on specific decoding methods or prior knowledge of the watermarking techniques. We develop WMD using foundations of offset learning, where a clean non-watermarked dataset enables us to isolate the influence of only watermarked samples in the reference dataset. Our comprehensive evaluations demonstrate the effectiveness of WMD, significantly outperforming naive detection methods, which only yield AUC scores around 0.5. In contrast, WMD consistently achieves impressive detection AUC scores, surpassing 0.9 in most single-watermark datasets and exceeding 0.7 in more challenging multi-watermark scenarios across diverse datasets and watermarking methods. As invisible watermarks become increasingly prevalent, while specific decoding techniques remain undisclosed, our approach provides a versatile solution and establishes a path toward increasing accountability, transparency, and trust in our digital visual content.

Peiyang Xu, Minzhou Pan, Zhaorun Chen et al.

With the rapid proliferation of digital media, the need for efficient and transparent safeguards against unsafe content is more critical than ever. Traditional image guardrail models, constrained by predefined categories, often misclassify content due to their pure feature-based learning without semantic reasoning. Moreover, these models struggle to adapt to emerging threats, requiring costly retraining for new threats. To address these limitations, we introduce SafeVision, a novel image guardrail that integrates human-like reasoning to enhance adaptability and transparency. Our approach incorporates an effective data collection and generation framework, a policy-following training pipeline, and a customized loss function. We also propose a diverse QA generation and training strategy to enhance learning effectiveness. SafeVision dynamically aligns with evolving safety policies at inference time, eliminating the need for retraining while ensuring precise risk assessments and explanations. Recognizing the limitations of existing unsafe image benchmarks, which either lack granularity or cover limited risks, we introduce VisionHarm, a high-quality dataset comprising two subsets: VisionHarm Third-party (VisionHarm-T) and VisionHarm Comprehensive(VisionHarm-C), spanning diverse harmful categories. Through extensive experiments, we show that SafeVision achieves state-of-the-art performance on different benchmarks. SafeVision outperforms GPT-4o by 8.6% on VisionHarm-T and by 15.5% on VisionHarm-C, while being over 16x faster. SafeVision sets a comprehensive, policy-following, and explainable image guardrail with dynamic adaptation to emerging threats.

Zhaorun Chen, Xun Liu, Mintong Kang et al.

As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world VLM vulnerabilities. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms into ARMs via model context protocol (MCP). To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack exploration algorithm. Extensive experiments on instance- and policy-based benchmarks show that ARMs achieves SOTA attack success rates, exceeding baselines by an average of 52.1% and surpassing 90% on Claude-4-Sonnet. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety dataset comprising over 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Safety fine-tuning with ARMs-Bench substantially improves the robustness of VLMs while preserving their general utility, providing actionable guidance to improve multimodal safety alignment against emerging threats.

Yi Zeng, Kevin Klyman, Andy Zhou et al.

We present a comprehensive AI risk taxonomy derived from eight government policies from the European Union, United States, and China and 16 company policies worldwide, making a significant step towards establishing a unified language for generative AI safety evaluation. We identify 314 unique risk categories organized into a four-tiered taxonomy. At the highest level, this taxonomy encompasses System & Operational Risks, Content Safety Risks, Societal Risks, and Legal & Rights Risks. The taxonomy establishes connections between various descriptions and approaches to risk, highlighting the overlaps and discrepancies between public and private sector conceptions of risk. By providing this unified framework, we aim to advance AI safety through information sharing across sectors and the promotion of best practices in risk mitigation for generative AI models and systems.

Minzhou Pan, Yi Zeng, Xue Lin et al.

In this study, we investigate the vulnerability of image watermarks to diffusion-model-based image editing, a challenge exacerbated by the computational cost of accessing gradient information and the closed-source nature of many diffusion models. To address this issue, we introduce JIGMARK. This first-of-its-kind watermarking technique enhances robustness through contrastive learning with pairs of images, processed and unprocessed by diffusion models, without needing a direct backpropagation of the diffusion process. Our evaluation reveals that JIGMARK significantly surpasses existing watermarking solutions in resilience to diffusion-model edits, demonstrating a True Positive Rate more than triple that of leading baselines at a 1% False Positive Rate while preserving image quality. At the same time, it consistently improves the robustness against other conventional perturbations (like JPEG, blurring, etc.) and malicious watermark attacks over the state-of-the-art, often by a large margin. Furthermore, we propose the Human Aligned Variation (HAV) score, a new metric that surpasses traditional similarity measures in quantifying the number of image derivatives from image editing.