AI-generated images (AIGIs) have reached a level of quality at which humans are unable to reliably distinguish them from real images. To counter the inherent risks of fraud and disinformation, their detection is a pressing challenge and an active research topic. While many of the presented methods claim high detection accuracy, they are often evaluated without considering their adversarial robustness. In real-world scenarios, malicious actors can modify AIGIs, e.g., through imperceptible perturbations, to evade detection systems. In this work, we provide a dataset to assess the robustness of AI-generated image detectors. We present RAID (Robust evaluation of AI-generated image Detectors), a benchmark dataset of 216k diverse and transferable adversarial images. The proposed dataset is created starting from images generated by four different text-to-image models (to promote diversity) and optimizing attacks against an ensemble of seven state-of-the-art detectors (to promote transferability). Extensive experiments show that our methodology generates adversarial images that transfer to unseen detectors, degrading the performance of every detector outside the attack ensemble that we evaluate, and provides a practical, worst-case robustness screen: failure on RAID is a strong indicator that a detector is not adversarially robust. We release our dataset at https://huggingface.co/datasets/aimagelab/RAID and evaluation code at https://github.com/pralab/RAID.
Despite advancements in Natural Language Processing (NLP), models remain vulnerable to adversarial attacks, such as synonym substitutions. While prior work has focused on improving robustness for feed-forward and convolutional architectures, the robustness of recurrent networks and modern state space models (SSMs), such as S4, remains understudied. These architectures pose unique challenges due to their sequential processing and complex parameter dynamics. In this paper, we introduce a novel regularization technique based on Growth Bound Matrices (GBM) to improve NLP model robustness by reducing the impact of input perturbations on model outputs. We focus on computing the GBM for three architectures: Long Short-Term Memory (LSTM), State Space models (S4), and Convolutional Neural Networks (CNN). Our method aims to (1) enhance resilience against word substitution attacks, (2) improve generalization on clean text, and (3) providing the first systematic analysis of SSM (S4) robustness. Extensive experiments across multiple architectures and benchmark datasets demonstrate that our method improves adversarial robustness by up to 8.8% over existing baselines. These results highlight the effectiveness of our approach, outperforming several state-of-the-art methods in adversarial defense. Codes are available at https://github.com/BouriMohammed/GBM
Fine-tuning has become the standard practice for adapting pre-trained (upstream) models to downstream tasks. However, the impact on model robustness is not well understood. In this work, we characterize the robustness-accuracy trade-off in fine-tuning. We evaluate the robustness and accuracy of fine-tuned models over 6 benchmark datasets and 7 different fine-tuning strategies. We observe a consistent trade-off between adversarial robustness and accuracy. Peripheral updates such as BitFit are more effective for simple tasks--over 75% above the average measured with area under the Pareto frontiers on CIFAR-10 and CIFAR-100. In contrast, fine-tuning information-heavy layers, such as attention layers via Compacter, achieves a better Pareto frontier on more complex tasks--57.5% and 34.6% above the average on Caltech-256 and CUB-200, respectively. Lastly, we observe that robustness of fine-tuning against out-of-distribution data closely tracks accuracy. These insights emphasize the need for robustness-aware fine-tuning to ensure reliable real-world deployments.
To guarantee safe and robust deployment of large language models (LLMs) at scale, it is critical to accurately assess their adversarial robustness. Existing adversarial attacks typically target harmful responses in single-point, greedy generations, overlooking the inherently stochastic nature of LLMs. In this paper, we propose a novel framework for adversarial robustness evaluation that explicitly models the entire output distribution, including tail-risks, providing better estimates for model robustness at scale. By casting the attack process as a resource allocation problem between optimization and sampling, we determine compute-optimal tradeoffs and show that integrating sampling into existing attacks boosts ASR by up to 48% and improves efficiency by up to two orders of magnitude. Our framework also enables us to analyze how different attack algorithms affect output harm distributions. Surprisingly, we find that most optimization strategies have little effect on output harmfulness. Finally, we introduce a data-free proof-of-concept objective based on entropy-maximization to demonstrate how our tail-aware perspective enables new optimization targets. Overall, our findings highlight the importance of tail-aware attacks and evaluation protocols to accurately assess and strengthen LLM safety.
Despite significant progress in designing powerful adversarial evasion attacks for robustness verification, the evaluation of these methods often remains inconsistent and unreliable. Many assessments rely on mismatched models, unverified implementations, and uneven computational budgets, which can lead to biased results and a false sense of security. Consequently, robustness claims built on such flawed testing protocols may be misleading and give a false sense of security. As a concrete step toward improving evaluation reliability, we present AttackBench, a benchmark framework developed to assess the effectiveness of gradient-based attacks under standardized and reproducible conditions. AttackBench serves as an evaluation tool that ranks existing attack implementations based on a novel optimality metric, which enables researchers and practitioners to identify the most reliable and effective attack for use in subsequent robustness evaluations. The framework enforces consistent testing conditions and enables continuous updates, making it a reliable foundation for robustness verification.
Existing defense methods fail to defend against unknown attacks and thus raise generalization issue of adversarial robustness. To remedy this problem, we attempt to delve into some underlying common characteristics among various attacks for generality. In this work, we reveal the commonly overlooked low entropy prior (LE) implied in various adversarial samples, and shed light on the universal robustness against unseen attacks in inference phase. LE prior is elaborated as two properties across various attacks as shown in Fig. 1 and Fig. 2: 1) low entropy misclassification for adversarial samples and 2) lower entropy prediction for higher attack intensity. This phenomenon stands in stark contrast to the naturally distributed samples. The LE prior can instruct existing test-time defense methods, thus we propose a two-stage REAL approach: Rectify Adversarial sample based on LE prior for test-time adversarial rectification. Specifically, to align adversarial samples more closely with clean samples, we propose to first rectify adversarial samples misclassified with low entropy by reverse maximizing prediction entropy, thereby eliminating their adversarial nature. To ensure the rectified samples can be correctly classified with low entropy, we carry out secondary rectification by forward minimizing prediction entropy, thus creating a Max-Min entropy optimization scheme. Further, based on the second property, we propose an attack-aware weighting mechanism to adaptively adjust the strengths of Max-Min entropy objectives. Experiments on several datasets show that REAL can greatly improve the performance of existing sample rectification models.
Reasoning Language Models (RLMs) have gained traction for their ability to perform complex, multi-step reasoning tasks through mechanisms such as Chain-of-Thought (CoT) prompting or fine-tuned reasoning traces. While these capabilities promise improved reliability, their impact on robustness to social biases remains unclear. In this work, we leverage the CLEAR-Bias benchmark, originally designed for Large Language Models (LLMs), to investigate the adversarial robustness of RLMs to bias elicitation. We systematically evaluate state-of-the-art RLMs across diverse sociocultural dimensions, using an LLM-as-a-judge approach for automated safety scoring and leveraging jailbreak techniques to assess the strength of built-in safety mechanisms. Our evaluation addresses three key questions: (i) how the introduction of reasoning capabilities affects model fairness and robustness; (ii) whether models fine-tuned for reasoning exhibit greater safety than those relying on CoT prompting at inference time; and (iii) how the success rate of jailbreak attacks targeting bias elicitation varies with the reasoning mechanisms employed. Our findings reveal a nuanced relationship between reasoning capabilities and bias safety. Surprisingly, models with explicit reasoning, whether via CoT prompting or fine-tuned reasoning traces, are generally more vulnerable to bias elicitation than base models without such mechanisms, suggesting reasoning may unintentionally open new pathways for stereotype reinforcement. Reasoning-enabled models appear somewhat safer than those relying on CoT prompting, which are particularly prone to contextual reframing attacks through storytelling prompts, fictional personas, or reward-shaped instructions. These results challenge the assumption that reasoning inherently improves robustness and underscore the need for more bias-aware approaches to reasoning design.
Adversarial robustness of neural networks is an increasingly important area of research, combining studies on computer vision models, large language models (LLMs), and others. With the release of JPEG AI -- the first standard for end-to-end neural image compression (NIC) methods -- the question of evaluating NIC robustness has become critically significant. However, previous research has been limited to a narrow range of codecs and attacks. To address this, we present \textbf{NIC-RobustBench}, the first open-source framework to evaluate NIC robustness and adversarial defenses' efficiency, in addition to comparing Rate-Distortion (RD) performance. The framework includes the largest number of codecs among all known NIC libraries and is easily scalable. The paper demonstrates a comprehensive overview of the NIC-RobustBench framework and employs it to analyze NIC robustness. Our code is available online at https://github.com/msu-video-group/NIC-RobustBench.
As large language models (LLMs) advance, concerns about their misconduct in complex social contexts intensify. Existing research overlooked the systematic understanding and assessment of their criminal capability in realistic interactions. We propose a unified framework PRISON, to quantify LLMs' criminal potential across five dimensions: False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement. Using structured crime scenarios adapted from classic films, we evaluate both criminal potential and anti-crime ability of LLMs via role-play. Results show that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements or evasion tactics, even without explicit instructions. Moreover, when placed in a detective role, models recognize deceptive behavior with only 41% accuracy on average, revealing a striking mismatch between conducting and detecting criminal behavior. These findings underscore the urgent need for adversarial robustness, behavioral alignment, and safety mechanisms before broader LLM deployment.
Adversarial examples have attracted significant attention over the years, yet understanding their frequency-based characteristics remains insufficient. In this paper, we investigate the intriguing properties of adversarial examples in the frequency domain for the image classification task, with the following key findings. (1) As the high-frequency components increase, the performance gap between adversarial and natural examples becomes increasingly pronounced. (2) The model performance against filtered adversarial examples initially increases to a peak and declines to its inherent robustness. (3) In Convolutional Neural Networks, mid- and high-frequency components of adversarial examples exhibit their attack capabilities, while in Transformers, low- and mid-frequency components of adversarial examples are particularly effective. These results suggest that different network architectures have different frequency preferences and that differences in frequency components between adversarial and natural examples may directly influence model robustness. Based on our findings, we further conclude with three useful proposals that serve as a valuable reference to the AI model security community.
Vision-Language Models (VLMs) such as CLIP have demonstrated remarkable capabilities in understanding relationships between visual and textual data through joint embedding spaces. Despite their effectiveness, these models remain vulnerable to adversarial attacks, particularly in the image modality, posing significant security concerns. Building upon our previous work on Adversarial Prompt Tuning (AdvPT), which introduced learnable text prompts to enhance adversarial robustness in VLMs without extensive parameter training, we present a significant extension by introducing the Neural Augmentor framework for Multi-modal Adversarial Prompt Tuning (NAP-Tuning).Our key innovations include: (1) extending AdvPT from text-only to multi-modal prompting across both text and visual modalities, (2) expanding from single-layer to multi-layer prompt architectures, and (3) proposing a novel architecture-level redesign through our Neural Augmentor approach, which implements feature purification to directly address the distortions introduced by adversarial attacks in feature space. Our NAP-Tuning approach incorporates token refiners that learn to reconstruct purified features through residual connections, allowing for modality-specific and layer-specific feature correction.Comprehensive experiments demonstrate that NAP-Tuning significantly outperforms existing methods across various datasets and attack types. Notably, our approach shows significant improvements over the strongest baselines under the challenging AutoAttack benchmark, outperforming them by 33.5% on ViT-B16 and 33.0% on ViT-B32 architectures while maintaining competitive clean accuracy.
The growing adoption of large pre-trained models in edge computing has made deploying model inference on mobile clients both practical and popular. These devices are inherently vulnerable to direct adversarial attacks, which pose a substantial threat to the robustness and security of deployed models. Federated adversarial training (FAT) has emerged as an effective solution to enhance model robustness while preserving client privacy. However, FAT frequently produces a generalized global model, which struggles to address the diverse and heterogeneous data distributions across clients, resulting in insufficiently personalized performance, while also encountering substantial communication challenges during the training process. In this paper, we propose \textit{Sylva}, a personalized collaborative adversarial training framework designed to deliver customized defense models for each client through a two-phase process. In Phase 1, \textit{Sylva} employs LoRA for local adversarial fine-tuning, enabling clients to personalize model robustness while drastically reducing communication costs by uploading only LoRA parameters during federated aggregation. In Phase 2, a game-based layer selection strategy is introduced to enhance accuracy on benign data, further refining the personalized model. This approach ensures that each client receives a tailored defense model that balances robustness and accuracy effectively. Extensive experiments on benchmark datasets demonstrate that \textit{Sylva} can achieve up to 50$\times$ improvements in communication efficiency compared to state-of-the-art algorithms, while achieving up to 29.5\% and 50.4\% enhancements in adversarial robustness and benign accuracy, respectively.
Conditional diffusion models (CDMs) have shown impressive performance across a range of generative tasks. Their ability to model the full data distribution has opened new avenues for analysis-by-synthesis in downstream discriminative learning. However, this same modeling capacity causes CDMs to entangle the class-defining features with irrelevant context, posing challenges to extracting robust and interpretable representations. To this end, we identify Canonical LAtent Representations (CLAReps), latent codes whose internal CDM features preserve essential categorical information while discarding non-discriminative signals. When decoded, CLAReps produce representative samples for each class, offering an interpretable and compact summary of the core class semantics with minimal irrelevant details. Exploiting CLAReps, we develop a novel diffusion-based feature-distillation paradigm, CaDistill. While the student has full access to the training set, the CDM as teacher transfers core class knowledge only via CLAReps, which amounts to merely 10 % of the training data in size. After training, the student achieves strong adversarial robustness and generalization ability, focusing more on the class signals instead of spurious background cues. Our findings suggest that CDMs can serve not just as image generators but also as compact, interpretable teachers that can drive robust representation learning.
Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarially robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of $53.2\%$ in AD and $68.5\%$ in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at https://github.com/rohban-lab/PatchGuard .
As deep learning models are increasingly deployed in high-risk applications, robust defenses against adversarial attacks and reliable performance guarantees become paramount. Moreover, accuracy alone does not provide sufficient assurance or reliable uncertainty estimates for these models. This study advances adversarial training by leveraging principles from Conformal Prediction. Specifically, we develop an adversarial attack method, termed OPSA (OPtimal Size Attack), designed to reduce the efficiency of conformal prediction at any significance level by maximizing model uncertainty without requiring coverage guarantees. Correspondingly, we introduce OPSA-AT (Adversarial Training), a defense strategy that integrates OPSA within a novel conformal training paradigm. Experimental evaluations demonstrate that our OPSA attack method induces greater uncertainty compared to baseline approaches for various defenses. Conversely, our OPSA-AT defensive model significantly enhances robustness not only against OPSA but also other adversarial attacks, and maintains reliable prediction. Our findings highlight the effectiveness of this integrated approach for developing trustworthy and resilient deep learning models for safety-critical domains. Our code is available at https://github.com/bjbbbb/Enhancing-Adversarial-Robustness-with-Conformal-Prediction.
Adversarial Robustness Distillation (ARD) has emerged as an effective method to enhance the robustness of lightweight deep neural networks against adversarial attacks. Current ARD approaches have leveraged a large robust teacher network to train one robust lightweight student. However, due to the diverse range of edge devices and resource constraints, current approaches require training a new student network from scratch to meet specific constraints, leading to substantial computational costs and increased CO2 emissions. This paper proposes Progressive Adversarial Robustness Distillation (ProARD), enabling the efficient one-time training of a dynamic network that supports a diverse range of accurate and robust student networks without requiring retraining. We first make a dynamic deep neural network based on dynamic layers by encompassing variations in width, depth, and expansion in each design stage to support a wide range of architectures. Then, we consider the student network with the largest size as the dynamic teacher network. ProARD trains this dynamic network using a weight-sharing mechanism to jointly optimize the dynamic teacher network and its internal student networks. However, due to the high computational cost of calculating exact gradients for all the students within the dynamic network, a sampling mechanism is required to select a subset of students. We show that random student sampling in each iteration fails to produce accurate and robust students.
Adversarial Training (AT) is widely recognized as an effective approach to enhance the adversarial robustness of Deep Neural Networks. As a variant of AT, Adversarial Robustness Distillation (ARD) has shown outstanding performance in enhancing the robustness of small models. However, both AT and ARD face robust fairness issue: these models tend to display strong adversarial robustness against some classes (easy classes) while demonstrating weak adversarial robustness against others (hard classes). This paper explores the underlying factors of this problem and points out the smoothness degree of soft labels for different classes significantly impacts the robust fairness from both empirical observation and theoretical analysis. Based on the above exploration, we propose Anti-Bias Soft Label Distillation (ABSLD) within the Knowledge Distillation framework to enhance the adversarial robust fairness. Specifically, ABSLD adaptively reduces the student's error risk gap between different classes, which is accomplished by adjusting the class-wise smoothness degree of teacher's soft labels during the training process, and the adjustment is managed by assigning varying temperatures to different classes. Additionally, as a label-based approach, ABSLD is highly adaptable and can be integrated with the sample-based methods. Extensive experiments demonstrate ABSLD outperforms state-of-the-art methods on the comprehensive performance of robustness and fairness.
While adversarial robustness and generalization have individually received substantial attention in the recent literature on quantum machine learning, their interplay is much less explored. In this chapter, we address this interplay for variational quantum models, which were recently proposed as function approximators in supervised learning. We discuss recent results quantifying both robustness and generalization via Lipschitz bounds, which explicitly depend on model parameters. Thus, they give rise to a regularization-based training approach for robust and generalizable quantum models, highlighting the importance of trainable data encoding strategies. The practical implications of the theoretical results are demonstrated with an application to time series analysis.
Adversarial training is among the most effective strategies for defending deep neural networks against adversarial examples. A key limitation of existing adversarial training approaches lies in their reliance on a fixed perturbation budget, which fails to account for instance-specific robustness characteristics. While prior works such as IAAT and MMA introduce instance-level adaptations, they often rely on heuristic or static approximations of data robustness. In this paper, we propose Dynamic Epsilon Scheduling (DES), a novel framework that adaptively adjusts the adversarial perturbation budget per instance and per training iteration. DES integrates three key factors: (1) the distance to the decision boundary approximated via gradient-based proxies, (2) prediction confidence derived from softmax entropy, and (3) model uncertainty estimated via Monte Carlo dropout. By combining these cues into a unified scheduling strategy, DES tailors the perturbation budget dynamically to guide more effective adversarial learning. Experimental results on CIFAR-10 and CIFAR-100 show that our method consistently improves both adversarial robustness and standard accuracy compared to fixed-epsilon baselines and prior adaptive methods. Moreover, we provide theoretical insights into the stability and convergence of our scheduling policy. This work opens a new avenue for instance-aware, data-driven adversarial training methods.
We introduce machine unlearning for speech tasks, a novel and underexplored research problem that aims to efficiently and effectively remove the influence of specific data from trained speech models without full retraining. This has important applications in privacy preservation, removal of outdated or noisy data, and bias mitigation. While machine unlearning has been studied in computer vision and natural language processing, its application to speech is largely unexplored due to the high-dimensional, sequential, and speaker-dependent nature of speech data. We define two fundamental speech unlearning tasks: sample unlearning, which removes individual data points (e.g., a voice recording), and class unlearning, which removes an entire category (e.g., all data from a speaker), while preserving performance on the remaining data. Experiments on keyword spotting and speaker identification demonstrate that unlearning speech data is significantly more challenging than unlearning image or text data. We conclude with key future directions in this area, including structured training, robust evaluation, feature-level unlearning, broader applications, scalable methods, and adversarial robustness.