The top two rows show easy cases, while the bottom three rows present hard cases, including crowdedness, complex backgrounds, and tiny objects.

The top two rows show easy cases, while the bottom three rows present hard cases, including crowdedness, complex backgrounds, and tiny objects.

Vision-Language Models (VLMs), built on pre-trained vision encoders and large language models (LLMs), have shown exceptional multi-modal understanding and dialog capabilities, positioning them as catalysts for the next technological revolution. However, while most VLM research focuses on enhancing multi-modal interaction, the risks of data misuse and leakage have been largely unexplored. This prompts the need for a comprehensive investigation of such risks in VLMs. In this paper, we conduct the first analysis of misuse and leakage detection in VLMs through the lens of membership inference attack (MIA). In specific, we focus on the instruction tuning data of VLMs, which is more likely to contain sensitive or unauthorized information. To address the limitation of existing MIA methods, we introduce a novel approach that infers membership based on a set of samples and their sensitivity to temperature, a unique parameter in VLMs. Based on this, we propose four membership inference methods, each tailored to different levels of background knowledge, ultimately arriving at the most challenging scenario. Our comprehensive evaluations show that these methods can accurately determine membership status, e.g., achieving an AUC greater than 0.8 targeting a small set consisting of only 5 samples on LLaVA.

Large Vision-Language Models (LVLMs) can reason effectively over both textual and visual inputs, but they tend to hallucinate syntactically coherent yet visually ungrounded contents. In this paper, we investigate the internal dynamics of hallucination by examining the tokens logits rankings throughout the generation process, revealing three key patterns in how LVLMs process information: (1) gradual visual information loss -- visually grounded tokens gradually become less favored throughout generation, and (2) early excitation -- semantically meaningful tokens achieve peak activation in the layers earlier than the final layer. (3) hidden genuine information -- visually grounded tokens though not being eventually decided still retain relatively high rankings at inference. Based on these insights, we propose VISTA (Visual Information Steering with Token-logit Augmentation), a training-free inference-time intervention framework that reduces hallucination while promoting genuine information. VISTA works by combining two complementary approaches: reinforcing visual information in activation space and leveraging early layer activations to promote semantically meaningful decoding. Compared to existing methods, VISTA requires no external supervision and is applicable to various decoding strategies. Extensive experiments show that VISTA on average reduces hallucination by abount 40% on evaluated open-ended generation task, and it consistently outperforms existing methods on four benchmarks across four architectures under three decoding strategies.

Understanding the vulnerabilities of Large Vision Language Models (LVLMs) to jailbreak attacks is essential for their responsible real-world deployment. Most previous work requires access to model gradients, or is based on human knowledge (prompt engineering) to complete jailbreak, and they hardly consider the interaction of images and text, resulting in inability to jailbreak in black box scenarios or poor performance. To overcome these limitations, we propose a Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for toxicity maximization, referred to as PBI-Attack. Our method begins by extracting malicious features from a harmful corpus using an alternative LVLM and embedding these features into a benign image as prior information. Subsequently, we enhance these features through bidirectional cross-modal interaction optimization, which iteratively optimizes the bimodal perturbations in an alternating manner through greedy search, aiming to maximize the toxicity of the generated response. The toxicity level is quantified using a well-trained evaluation model. Experiments demonstrate that PBI-Attack outperforms previous state-of-the-art jailbreak methods, achieving an average attack success rate of 92.5% across three open-source LVLMs and around 67.3% on three closed-source LVLMs. Disclaimer: This paper contains potentially disturbing and offensive content.

Jailbreaking methods, which induce Multi-modal Large Language Models (MLLMs) to output harmful responses, raise significant safety concerns. Among these methods, gradient-based approaches, which use gradients to generate malicious prompts, have been widely studied due to their high success rates in white-box settings, where full access to the model is available. However, these methods have notable limitations: they require white-box access, which is not always feasible, and involve high memory usage. To address scenarios where white-box access is unavailable, attackers often resort to transfer attacks. In transfer attacks, malicious inputs generated using white-box models are applied to black-box models, but this typically results in reduced attack performance. To overcome these challenges, we propose Zer0-Jack, a method that bypasses the need for white-box access by leveraging zeroth-order optimization. We propose patch coordinate descent to efficiently generate malicious image inputs to directly attack black-box MLLMs, which significantly reduces memory usage further. Through extensive experiments, Zer0-Jack achieves a high attack success rate across various models, surpassing previous transfer-based methods and performing comparably with existing white-box jailbreak techniques. Notably, Zer0-Jack achieves a 95\% attack success rate on MiniGPT-4 with the Harmful Behaviors Multi-modal Dataset on a black-box setting, demonstrating its effectiveness. Additionally, we show that Zer0-Jack can directly attack commercial MLLMs such as GPT-4o. Codes are provided in the supplement.

Multimodal large language models (MLLMs) have revolutionized vision-language understanding but are vulnerable to multimodal jailbreak attacks, where adversaries meticulously craft inputs to elicit harmful or inappropriate responses. We propose UniGuard, a novel multimodal safety guardrail that jointly considers the unimodal and cross-modal harmful signals. UniGuard is trained such that the likelihood of generating harmful responses in a toxic corpus is minimized, and can be seamlessly applied to any input prompt during inference with minimal computational costs. Extensive experiments demonstrate the generalizability of UniGuard across multiple modalities and attack strategies. It demonstrates impressive generalizability across multiple state-of-the-art MLLMs, including LLaVA, Gemini Pro, GPT-4, MiniGPT-4, and InstructBLIP, thereby broadening the scope of our solution.

Visual Language Models (VLMs) are vulnerable to adversarial attacks, especially those from adversarial images, which is however under-explored in literature. To facilitate research on this critical safety problem, we first construct a new laRge-scale Adervsarial images dataset with Diverse hArmful Responses (RADAR), given that existing datasets are either small-scale or only contain limited types of harmful responses. With the new RADAR dataset, we further develop a novel and effective iN-time Embedding-based AdveRSarial Image DEtection (NEARSIDE) method, which exploits a single vector that distilled from the hidden states of VLMs, which we call the attacking direction, to achieve the detection of adversarial images against benign ones in the input. Extensive experiments with two victim VLMs, LLaVA and MiniGPT-4, well demonstrate the effectiveness, efficiency, and cross-model transferrability of our proposed method. Our code is available at https://github.com/mob-scu/RADAR-NEARSIDE

Despite their superb multimodal capabilities, Vision-Language Models (VLMs) have been shown to be vulnerable to jailbreak attacks, which are inference-time attacks that induce the model to output harmful responses with tricky prompts. It is thus essential to defend VLMs against potential jailbreaks for their trustworthy deployment in real-world applications. In this work, we focus on black-box defense for VLMs against jailbreak attacks. Existing black-box defense methods are either unimodal or bimodal. Unimodal methods enhance either the vision or language module of the VLM, while bimodal methods robustify the model through text-image representation realignment. However, these methods suffer from two limitations: 1) they fail to fully exploit the cross-modal information, or 2) they degrade the model performance on benign inputs. To address these limitations, we propose a novel blue-team method BlueSuffix that defends the black-box target VLM against jailbreak attacks without compromising its performance. BlueSuffix includes three key components: 1) a visual purifier against jailbreak images, 2) a textual purifier against jailbreak texts, and 3) a blue-team suffix generator finetuned via reinforcement learning for enhancing cross-modal robustness. We empirically show on three VLMs (LLaVA, MiniGPT-4, and Gemini) and two safety benchmarks (MM-SafetyBench and RedTeam-2K) that BlueSuffix outperforms the baseline defenses by a significant margin. Our BlueSuffix opens up a promising direction for defending VLMs against jailbreak attacks. There has been a notable surge in research focusing on incorporating multimodal capabilities into Large Language Models (LLMs), leading to the emergence of Vision-Language Models (VLMs), such as OpenAI's GPT-4o (Achiam et al., 2023) and Google's Gemini 1.5 (Reid et al., 2024).

Large vision-language models (LVLMs) integrate visual information into large language models, showcasing remarkable multi-modal conversational capabilities. However, the visual modules introduces new challenges in terms of robustness for LVLMs, as attackers can craft adversarial images that are visually clean but may mislead the model to generate incorrect answers. In general, LVLMs rely on vision encoders to transform images into visual tokens, which are crucial for the language models to perceive image contents effectively. Therefore, we are curious about one question: Can LVLMs still generate correct responses when the encoded visual tokens are attacked and disrupting the visual information? To this end, we propose a non-targeted attack method referred to as VT-Attack (Visual Tokens Attack), which constructs adversarial examples from multiple perspectives, with the goal of comprehensively disrupting feature representations and inherent relationships as well as the semantic properties of visual tokens output by image encoders. Using only access to the image encoder in the proposed attack, the generated adversarial examples exhibit transferability across diverse LVLMs utilizing the same image encoder and generality across different tasks. Extensive experiments validate the superior attack performance of the VT-Attack over baseline methods, demonstrating its effectiveness in attacking LVLMs with image encoders, which in turn can provide guidance on the robustness of LVLMs, particularly in terms of the stability of the visual feature space.