Feature-Space Adversarial Robustness Certification for Multimodal Large Language Models

Multimodal large language models (MLLMs) exhibit strong capabilities across diverse applications, yet remain vulnerable to adversarial perturbations that distort their feature representations and induce erroneous predictions. To address this vulnerability, we propose Feature-space Smoothing (FS), a general framework that provides certified robustness guarantees at the feature representation level of MLLMs. We theoretically prove that FS converts a given feature extractor into a smoothed variant that is guaranteed a certified lower bound on the cosine similarity between clean and adversarial features under $\ell_2$-bounded perturbations. Moreover, we establish that the value of this Feature Cosine Similarity Bound (FCSB) is determined by the intrinsic Gaussian robustness score of the given encoder. Building on this insight, we introduce the Gaussian Smoothness Booster (GSB), a plug-and-play module that enhances the Gaussian robustness score of pretrained MLLMs, thereby strengthening the robustness guaranteed by FS, without requiring additional MLLM retraining. Extensive experiments demonstrate that applying the FS to various MLLMs yields strong certified feature-space robustness and consistently leads to robust task-oriented performance across diverse applications.

Provable Robustness in Multimodal Large Language Models via Feature Space Smoothing

Multimodal large language models (MLLMs) exhibit strong capabilities across diverse applications, yet remain vulnerable to adversarial perturbations that distort their feature representations and induce erroneous predictions. To address this vulnerability, we propose the Feature-space Smoothing (FS) and theoretically prove that FS offers certified robustness on the feature representations of MLLMs. Specifically, FS transforms any feature encoder into a smoothed variant that is guaranteed to maintain a certified lower bound on the feature cosine similarity between clean and adversarial representations under $\ell_2$-bounded attacks. Moreover, we indicate that the value of this Feature Cosine Similarity Bound (FCSB) derived from FS can be improved by enlarging the defined Gaussian robustness score on the vanilla encoder. Building upon this, we introduce the Purifier and Smoothness Mapper (PSM), a plug-and-play module that improves the Gaussian robustness score of MLLMs and thus enhances their certified robustness under FS, without requiring any retraining on MLLMs. We demonstrate that the FS with PSM not only provides a strong theoretical robustness guarantee but also exhibits superior empirical performance compared to adversarial training. Extensive experiments across diverse MLLMs and downstream tasks indicate the effectiveness of the FS-PSM, reducing the Attack Success Rate (ASR) of various white-box attacks from nearly 90\% to about 1\%.

From Pretrain to Pain: Adversarial Vulnerability of Video Foundation Models Without Task Knowledge

Large-scale Video Foundation Models (VFMs) has significantly advanced various video-related tasks, either through task-specific models or Multi-modal Large Language Models (MLLMs). However, the open accessibility of VFMs also introduces critical security risks, as adversaries can exploit full knowledge of the VFMs to launch potent attacks. This paper investigates a novel and practical adversarial threat scenario: attacking downstream models or MLLMs fine-tuned from open-source VFMs, without requiring access to the victim task, training data, model query, and architecture. In contrast to conventional transfer-based attacks that rely on task-aligned surrogate models, we demonstrate that adversarial vulnerabilities can be exploited directly from the VFMs. To this end, we propose the Transferable Video Attack (TVA), a temporal-aware adversarial attack method that leverages the temporal representation dynamics of VFMs to craft effective perturbations. TVA integrates a bidirectional contrastive learning mechanism to maximize the discrepancy between the clean and adversarial features, and introduces a temporal consistency loss that exploits motion cues to enhance the sequential impact of perturbations. TVA avoids the need to train expensive surrogate models or access to domain-specific data, thereby offering a more practical and efficient attack strategy. Extensive experiments across 24 video-related tasks demonstrate the efficacy of TVA against downstream models and MLLMs, revealing a previously underexplored security vulnerability in the deployment of video models.

MTL-UE: Learning to Learn Nothing for Multi-Task Learning

Most existing unlearnable strategies focus on preventing unauthorized users from training single-task learning (STL) models with personal data. Nevertheless, the paradigm has recently shifted towards multi-task data and multi-task learning (MTL), targeting generalist and foundation models that can handle multiple tasks simultaneously. Despite their growing importance, MTL data and models have been largely neglected while pursuing unlearnable strategies. This paper presents MTL-UE, the first unified framework for generating unlearnable examples for multi-task data and MTL models. Instead of optimizing perturbations for each sample, we design a generator-based structure that introduces label priors and class-wise feature embeddings which leads to much better attacking performance. In addition, MTL-UE incorporates intra-task and inter-task embedding regularization to increase inter-class separation and suppress intra-class variance which enhances the attack robustness greatly. Furthermore, MTL-UE is versatile with good supports for dense prediction tasks in MTL. It is also plug-and-play allowing integrating existing surrogate-dependent unlearnable methods with little adaptation. Extensive experiments show that MTL-UE achieves superior attacking performance consistently across 4 MTL datasets, 3 base UE methods, 5 model backbones, and 5 MTL task-weighting strategies.

Open-set Anomaly Segmentation in Complex Scenarios

Precise segmentation of out-of-distribution (OoD) objects, herein referred to as anomalies, is crucial for the reliable deployment of semantic segmentation models in open-set, safety-critical applications, such as autonomous driving. Current anomalous segmentation benchmarks predominantly focus on favorable weather conditions, resulting in untrustworthy evaluations that overlook the risks posed by diverse meteorological conditions in open-set environments, such as low illumination, dense fog, and heavy rain. To bridge this gap, this paper introduces the ComsAmy, a challenging benchmark specifically designed for open-set anomaly segmentation in complex scenarios. ComsAmy encompasses a wide spectrum of adverse weather conditions, dynamic driving environments, and diverse anomaly types to comprehensively evaluate the model performance in realistic open-world scenarios. Our extensive evaluation of several state-of-the-art anomalous segmentation models reveals that existing methods demonstrate significant deficiencies in such challenging scenarios, highlighting their serious safety risks for real-world deployment. To solve that, we propose a novel energy-entropy learning (EEL) strategy that integrates the complementary information from energy and entropy to bolster the robustness of anomaly segmentation under complex open-world environments. Additionally, a diffusion-based anomalous training data synthesizer is proposed to generate diverse and high-quality anomalous images to enhance the existing copy-paste training data synthesizer. Extensive experimental results on both public and ComsAmy benchmarks demonstrate that our proposed diffusion-based synthesizer with energy and entropy learning (DiffEEL) serves as an effective and generalizable plug-and-play method to enhance existing models, yielding an average improvement of around 4.96% in $\rm{AUPRC}$ and 9.87% in $\rm{FPR}_{95}$.

Towards Model Resistant to Transferable Adversarial Examples via Trigger Activation

Adversarial examples, characterized by imperceptible perturbations, pose significant threats to deep neural networks by misleading their predictions. A critical aspect of these examples is their transferability, allowing them to deceive {unseen} models in black-box scenarios. Despite the widespread exploration of defense methods, including those on transferability, they show limitations: inefficient deployment, ineffective defense, and degraded performance on clean images. In this work, we introduce a novel training paradigm aimed at enhancing robustness against transferable adversarial examples (TAEs) in a more efficient and effective way. We propose a model that exhibits random guessing behavior when presented with clean data $\boldsymbol{x}$ as input, and generates accurate predictions when with triggered data $\boldsymbol{x}+\boldsymbol{\tau}$. Importantly, the trigger $\boldsymbol{\tau}$ remains constant for all data instances. We refer to these models as \textbf{models with trigger activation}. We are surprised to find that these models exhibit certain robustness against TAEs. Through the consideration of first-order gradients, we provide a theoretical analysis of this robustness. Moreover, through the joint optimization of the learnable trigger and the model, we achieve improved robustness to transferable attacks. Extensive experiments conducted across diverse datasets, evaluating a variety of attacking methods, underscore the effectiveness and superiority of our approach.

Backdoor Attacks against No-Reference Image Quality Assessment Models via A Scalable Trigger

No-Reference Image Quality Assessment (NR-IQA), responsible for assessing the quality of a single input image without using any reference, plays a critical role in evaluating and optimizing computer vision systems, e.g., low-light enhancement. Recent research indicates that NR-IQA models are susceptible to adversarial attacks, which can significantly alter predicted scores with visually imperceptible perturbations. Despite revealing vulnerabilities, these attack methods have limitations, including high computational demands, untargeted manipulation, limited practical utility in white-box scenarios, and reduced effectiveness in black-box scenarios. To address these challenges, we shift our focus to another significant threat and present a novel poisoning-based backdoor attack against NR-IQA (BAIQA), allowing the attacker to manipulate the IQA model's output to any desired target value by simply adjusting a scaling coefficient $\alpha$ for the trigger. We propose to inject the trigger in the discrete cosine transform (DCT) domain to improve the local invariance of the trigger for countering trigger diminishment in NR-IQA models due to widely adopted data augmentations. Furthermore, the universal adversarial perturbations (UAP) in the DCT space are designed as the trigger, to increase IQA model susceptibility to manipulation and improve attack effectiveness. In addition to the heuristic method for poison-label BAIQA (P-BAIQA), we explore the design of clean-label BAIQA (C-BAIQA), focusing on $\alpha$ sampling and image data refinement, driven by theoretical insights we reveal. Extensive experiments on diverse datasets and various NR-IQA models demonstrate the effectiveness of our attacks. Code will be released at https://github.com/yuyi-sd/BAIQA.

Transferable Adversarial Attacks on SAM and Its Downstream Models

The utilization of large foundational models has a dilemma: while fine-tuning downstream tasks from them holds promise for making use of the well-generalized knowledge in practical applications, their open accessibility also poses threats of adverse usage. This paper, for the first time, explores the feasibility of adversarial attacking various downstream models fine-tuned from the segment anything model (SAM), by solely utilizing the information from the open-sourced SAM. In contrast to prevailing transfer-based adversarial attacks, we demonstrate the existence of adversarial dangers even without accessing the downstream task and dataset to train a similar surrogate model. To enhance the effectiveness of the adversarial attack towards models fine-tuned on unknown datasets, we propose a universal meta-initialization (UMI) algorithm to extract the intrinsic vulnerability inherent in the foundation model, which is then utilized as the prior knowledge to guide the generation of adversarial perturbations. Moreover, by formulating the gradient difference in the attacking process between the open-sourced SAM and its fine-tuned downstream models, we theoretically demonstrate that a deviation occurs in the adversarial update direction by directly maximizing the distance of encoded feature embeddings in the open-sourced SAM. Consequently, we propose a gradient robust loss that simulates the associated uncertainty with gradient-based noise augmentation to enhance the robustness of generated adversarial examples (AEs) towards this deviation, thus improving the transferability. Extensive experiments demonstrate the effectiveness of the proposed universal meta-initialized and gradient robust adversarial attack (UMI-GRAT) toward SAMs and their downstream models. Code is available at https://github.com/xiasong0501/GRAT.

Purify Unlearnable Examples via Rate-Constrained Variational Autoencoders

Unlearnable examples (UEs) seek to maximize testing error by making subtle modifications to training examples that are correctly labeled. Defenses against these poisoning attacks can be categorized based on whether specific interventions are adopted during training. The first approach is training-time defense, such as adversarial training, which can mitigate poisoning effects but is computationally intensive. The other approach is pre-training purification, e.g., image short squeezing, which consists of several simple compressions but often encounters challenges in dealing with various UEs. Our work provides a novel disentanglement mechanism to build an efficient pre-training purification method. Firstly, we uncover rate-constrained variational autoencoders (VAEs), demonstrating a clear tendency to suppress the perturbations in UEs. We subsequently conduct a theoretical analysis for this phenomenon. Building upon these insights, we introduce a disentangle variational autoencoder (D-VAE), capable of disentangling the perturbations with learnable class-wise embeddings. Based on this network, a two-stage purification approach is naturally developed. The first stage focuses on roughly eliminating perturbations, while the second stage produces refined, poison-free results, ensuring effectiveness and robustness across various scenarios. Extensive experiments demonstrate the remarkable performance of our method across CIFAR-10, CIFAR-100, and a 100-class ImageNet-subset. Code is available at https://github.com/yuyi-sd/D-VAE.

Mitigating the Curse of Dimensionality for Certified Robustness via Dual Randomized Smoothing

Randomized Smoothing (RS) has been proven a promising method for endowing an arbitrary image classifier with certified robustness. However, the substantial uncertainty inherent in the high-dimensional isotropic Gaussian noise imposes the curse of dimensionality on RS. Specifically, the upper bound of ${\ell_2}$ certified robustness radius provided by RS exhibits a diminishing trend with the expansion of the input dimension $d$, proportionally decreasing at a rate of $1/\sqrt{d}$. This paper explores the feasibility of providing ${\ell_2}$ certified robustness for high-dimensional input through the utilization of dual smoothing in the lower-dimensional space. The proposed Dual Randomized Smoothing (DRS) down-samples the input image into two sub-images and smooths the two sub-images in lower dimensions. Theoretically, we prove that DRS guarantees a tight ${\ell_2}$ certified robustness radius for the original input and reveal that DRS attains a superior upper bound on the ${\ell_2}$ robustness radius, which decreases proportionally at a rate of $(1/\sqrt m + 1/\sqrt n )$ with $m+n=d$. Extensive experiments demonstrate the generalizability and effectiveness of DRS, which exhibits a notable capability to integrate with established methodologies, yielding substantial improvements in both accuracy and ${\ell_2}$ certified robustness baselines of RS on the CIFAR-10 and ImageNet datasets. Code is available at https://github.com/xiasong0501/DRS.