Adversarial Examples

Current safety evaluations of large language models (LLMs) create a dangerous illusion of universal protection by aggregating harms under generic categories such as "Identity Hate", obscuring vulnerabilities toward specific populations. In this work, we expose the Selective Safety Trap: a systemic failure mode where models robustly defend specific populations while leaving underrepresented communities highly vulnerable to identical adversarial attacks. To systematically audit this phenomenon, we introduce MiJaBench, a bilingual (English-Portuguese) adversarial benchmark comprising 43,961 controlled jailbreaking prompts across 16 minority groups. By evaluating 14 state-of-the-art LLMs on MiJaBench, we curate 615,454 prompt-response pairs that compose MiJaBench-Align, revealing that safety alignment is not a uniform semantic capability but a demographic hierarchy, with defense rates fluctuating by up to 42% within the same model solely based on the target group. This disparity persists across architectures and languages and is amplified by scaling, indicating that current alignment methods learn group-specific safeguards rather than a generalized notion of harm. Through targeted direct preference optimization (DPO) on a 1B-parameter baseline, we achieve strong zero-shot safety generalizations to entirely unseen demographics and complex attack strategies. We release all datasets and scripts to provide the community with a concrete pathway toward equitable, transferable safety alignment.

Vision-language models (VLMs) are increasingly integrated into clinical workflows. However, existing benchmarks primarily assess performance on common anatomical presentations and fail to capture the challenges posed by rare variants. To address this gap, we introduce AdversarialAnatomyBench, the first benchmark comprising naturally occurring rare anatomical variants across diverse imaging modalities and anatomical regions. We call such variants that violate learned priors about "typical" human anatomy natural adversarial anatomy. Benchmarking 25 state-of-the-art VLMs with AdversarialAnatomyBench yielded three key insights. First, when queried with basic medical perception tasks, mean accuracy dropped from 71% on typical to 28% on atypical anatomy. Even the best-performing models, GPT-5, Gemini 2.5 Pro, and Llama 4 Maverick, showed performance drops of 41-51%. Second, model errors closely mirrored expected anatomical biases. Third, neither model scaling nor interventions, including bias-aware prompting and test-time reasoning, resolved these issues. These findings highlight a critical limitation in current VLMs: their poor generalization to rare anatomical presentations. AdversarialAnatomyBench provides a foundation for systematically measuring and mitigating anatomical bias in multimodal medical artificial intelligence (AI) systems.

Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods. The code is available at https://github.com/tairanhuang/seba online.

Deep learning models deployed in safety critical applications like autonomous driving use simulations to test their robustness against adversarial attacks in realistic conditions. However, these simulations are non-differentiable, forcing researchers to create attacks that do not integrate simulation environmental factors, reducing attack success. To address this limitation, we introduce UNDREAM, the first software framework that bridges the gap between photorealistic simulators and differentiable renderers to enable end-to-end optimization of adversarial perturbations on any 3D objects. UNDREAM enables manipulation of the environment by offering complete control over weather, lighting, backgrounds, camera angles, trajectories, and realistic human and object movements, thereby allowing the creation of diverse scenes. We showcase a wide array of distinct physically plausible adversarial objects that UNDREAM enables researchers to swiftly explore in different configurable environments. This combination of photorealistic simulation and differentiable optimization opens new avenues for advancing research of physical adversarial attacks.

Vision Language Models (VLMs) are increasingly being used in a broad range of applications, bringing their security and behavioral control to the forefront. While existing approaches for behavioral control or output redirection, like system prompting in VLMs, are easily detectable and often ineffective, activation-based steering vectors require invasive runtime access to model internals--incompatible with API-based services and closed-source deployments. We introduce VISOR (Visual Input-based Steering for Output Redirection), a novel method that achieves sophisticated behavioral control through optimized visual inputs alone. By crafting universal steering images that induce target activation patterns, VISOR enables practical deployment across all VLM serving modalities while remaining imperceptible compared to explicit textual instructions. We validate VISOR on LLaVA-1.5-7B across three critical alignment tasks: refusal, sycophancy and survival instinct. A single 150KB steering image matches steering vector performance within 1-2% for positive behavioral shifts while dramatically exceeding it for negative steering--achieving up to 25% shifts from baseline compared to steering vectors' modest changes. Unlike system prompting (3-4% shifts), VISOR provides robust bidirectional control while maintaining 99.9% performance on 14,000 unrelated MMLU tasks. Beyond eliminating runtime overhead and model access requirements, VISOR exposes a critical security vulnerability: adversaries can achieve sophisticated behavioral manipulation through visual channels alone, bypassing text-based defenses. Our work fundamentally re-imagines multimodal model control and highlights the urgent need for defenses against visual steering attacks.

What fundamentally distinguishes an adversarial attack from a misclassification due to limited model expressivity or finite data? In this work, we investigate this question in the setting of high-dimensional binary classification, where statistical effects due to limited data availability play a central role. We introduce a new error metric that precisely capture this distinction, quantifying model vulnerability to consistent adversarial attacks -- perturbations that preserve the ground-truth labels. Our main technical contribution is an exact and rigorous asymptotic characterization of these metrics in both well-specified models and latent space models, revealing different vulnerability patterns compared to standard robust error measures. The theoretical results demonstrate that as models become more overparameterized, their vulnerability to label-preserving perturbations grows, offering theoretical insight into the mechanisms underlying model sensitivity to adversarial attacks.

Deep neural networks (DNNs) are highly susceptible to adversarial examples---small, malicious perturbations that can cause incorrect predictions. We introduce a lightweight, plug-in detector that uses internal layer-wise inconsistencies within the target model and requires only benign data for fitting and calibration. The approach is motivated by the A Few Large Shifts Assumption, an empirical hypothesis that adversarial perturbations often produce large, localized growth in representation changes across a small number of consecutive layers, connecting adversarial behavior to layer-wise Lipschitz continuity. We develop two complementary scores---Recovery Testing (RT) for intermediate-layer inconsistency and Logit-layer Testing (LT) for augmentation-induced output instability---and fuse them through RLT. Across CIFAR-10, CIFAR-100, and ImageNet, RLT achieves strong detection performance under standard attacks with substantially lower overhead than detector families requiring external encoders or reference-set retrieval. We further study its behavior under adaptive attacks, at low false-positive operating points, and under benign distribution shifts. The code is available here: https://github.com/c0510gy/AFLS-AED.

In this paper, we study nonconvex constrained stochastic zeroth-order optimization problems with exact constraints and stochastic objective evaluations. To solve this class of problems, we propose a framework of mirror descent linearized augmented Lagrangian methods that employs two-point stochastic zeroth-order gradient estimators and exploits non-Euclidean mirror descent geometry. Under mild assumptions, we establish oracle complexity guarantees for finding an $ε$-KKT point parameterized by $p \geq 2$. Under Rademacher smoothing, our analysis reveals a trade-off between the variance of the zeroth-order gradient estimators and the smoothness of the mirror map. In the high-accuracy regime, the resulting effective oracle complexity is $\mathcal{O}(p d^{2/p}ε^{-3})$ for $p \in [2,2\ln d]$ and $\mathcal{O}(\ln d\,ε^{-3})$ for $p > 2\ln d$. These bounds reduce the dimension dependence in the leading term. When $p=2$, our method recovers the Euclidean setting with an oracle complexity of $\mathcal{O}(dε^{-3})$, improving the $ε$-dependence over existing methods. Furthermore, to eliminate initial near-feasibility requirements, we introduce a multi-stage scheme that finds an $ε$-KKT point within $\mathcal{O}(1+\log\log(e/ε))$ stages while maintaining the leading-order complexity. Numerical tests on QCQPs, black-box adversarial attacks, and fairness-constrained classification demonstrate the effectiveness of our proposed method.

Machine learning systems deployed in distributed or federated environments are highly susceptible to adversarial manipulations, particularly availability attacks -- rendering the trained model unavailable. Prior research in distributed ML has demonstrated such adversarial effects through the injection of gradients or data poisoning. In this work, we ask whether comparable degradation is still possible under a substantially more constrained action space: the adversary may only flip a limited number of labels of existing training examples, without modifying features, injecting samples, or directly controlling gradients. We analyze the extent of damage caused by constrained label flipping attacks against distributed learning under mean aggregation -- the dominant baseline in research and production. Focusing on classification problems, (1) we propose a novel formalization of label flipping attacks as a per-round constrained optimization problem, derive a greedy label-selection rule for logistic regression, and empirically evaluate it beyond its derivation setting, including on MLPs and robust aggregators. The rule is provably per-epoch optimal for the attacker under the mean aggregator. (2) Empirically, we show that optimized label flipping can cause substantial accuracy degradation while outperforming random label flipping under similar budgets. (3) We shed light on an interesting interplay between what the attacker gains from more write-access versus what they gain from more flipping budget. (4) Finally, although the attack is derived for mean aggregation, we find that it can transfer empirically to the coordinate-wise median and trimmed mean aggregators, where its effectiveness approaches that of the Little-is-Enough gradient attack. This demonstrates that even highly constrained label-flipping adversaries can pose a significant availability threat to distributed learning.

Large Vision-Language Models (LVLMs) have achieved remarkable performance on multimodal tasks but remain highly vulnerable to small adversarial perturbations in input images. Existing attacks typically target the vision encoder's final output embeddings, implicitly treating the encoder as a uniform attack surface, while a systematic analysis of which internal components are most vulnerable has remained largely unexplored. We show such analysis is essential, as adversarial vulnerability in LVLM vision encoders is structurally concentrated rather than uniformly distributed. Building on this, we propose Vision Encoder Vulnerable-Component-Targeted Universal Adversarial Perturbation (VEV-UAP), a task-agnostic and cost-efficient attack framework. Through a component- and layer-wise analysis of attention mechanisms, we identify the value components in middle layers as critical vulnerabilities that strongly influence downstream language model behavior. VEV-UAP selectively targets these components to generate a single universal perturbation shared across images, without involving textual inputs or the language model during optimization. Experiments across multiple LVLMs and tasks show VEV-UAP achieves state-of-the-art attack success rates with reduced computational overhead. Moreover, a single VEV-UAP transfers across LVLMs sharing the same vision encoder, even when paired with different language models, making it a practical framework for scalable robustness evaluation.

Deep neural networks are highly vulnerable to adversarial perturbations, raising serious safety concerns in the real-world systems. While prior work mainly explores single-task attacks or jointly degrading all tasks in multi-task models, practical scenarios often demand more selective and stealthy attack strategies. To address this challenge, we propose Stealthy Multi-Task Adversarial Attack (SMTA$^{2}$), a novel framework that selectively degrades a targeted task while strictly preserving the performance of non-targeted tasks. We formulate this objective as a constrained multi-objective optimization problem and design task-aware adversarial perturbations that maximize degradation on the targeted task without causing collateral damage on non-targeted tasks. To enhance practicality, we further introduce an automated loss-weight tuning strategy that dynamically balances attack and preservation objectives. Experiments on two multi-task benchmarks NYUv2 and Cityscapes demonstrate that SMTA$^{2}$ achieves strong attack performance on targeted tasks while maintaining non-targeted tasks intact on both undefended and adversarially trained models, establishing the first systematic framework for stealthy and selective multi-task attack framework.

The rapid advancement of Generative Artificial Intelligence (GenAI) capabilities is accompanied by a concerning rise in its misuse. In particular the generation of credible misinformation in the form of images poses a significant threat to the public trust in democratic processes. Consequently, there is an urgent need to develop tools to reliably distinguish between authentic and AI-generated content. The majority of detection methods are based on neural networks that are trained to recognize forensic artifacts. In this work, we demonstrate that current state-of-the-art classifiers are vulnerable to adversarial examples under real-world conditions. Through extensive experiments, comprising four detection methods and five attack algorithms, we show that an attacker can dramatically decrease classification performance, without internal knowledge of the detector's architecture. Notably, most attacks remain effective even when images are degraded during the upload to, e.g., social media platforms. In a case study, we demonstrate that these robustness challenges are also found in commercial tools by conducting black-box attacks on HIVE, a proprietary online GenAI media detector. In addition, we evaluate the robustness of using generated features of a robust pre-trained model and showed that this increases the robustness, while not reaching the performance on benign inputs. These results, along with the increasing potential of GenAI to erode public trust, underscore the need for more research and new perspectives on methods to prevent its misuse.

Voice Authentication Systems (VAS) use unique vocal characteristics for verification. They are increasingly integrated into high-security sectors such as banking and healthcare. Despite their improvements using deep learning, they face severe vulnerabilities from sophisticated threats like deepfakes and adversarial attacks. The emergence of realistic voice cloning complicates detection, as systems struggle to distinguish authentic from synthetic audio. While anti-spoofing countermeasures (CMs) exist to mitigate these risks, many rely on static detection models that can be bypassed by novel adversarial methods, leaving a critical security gap. To demonstrate this vulnerability, we propose the Spectral Masking and Interpolation Attack (SMIA), a novel method that strategically manipulates inaudible frequency regions of AI-generated audio. By altering the voice in imperceptible zones to the human ear, SMIA creates adversarial samples that sound authentic while deceiving CMs. We conducted a comprehensive evaluation of our attack against state-of-the-art (SOTA) models across multiple tasks, under simulated real-world conditions. SMIA achieved a strong attack success rate (ASR) of at least 82% against combined VAS/CM systems, at least 97.5% against standalone speaker verification systems, and 100% against countermeasures. These findings conclusively demonstrate that current security postures are insufficient against adaptive adversarial attacks. This work highlights the urgent need for a paradigm shift toward next-generation defenses that employ dynamic, context-aware frameworks capable of evolving with the threat landscape.