Adversarial training has emerged as a highly effective way to improve the robustness of deep neural networks (DNNs). It is typically conceptualized as a min-max optimization problem over model weights and adversarial perturbations, where the weights are optimized using gradient descent methods, such as SGD. In this paper, we propose a novel approach by treating model weights as random variables, which paves the way for enhancing adversarial training through \textbf{S}econd-Order \textbf{S}tatistics \textbf{O}ptimization (S$^2$O) over model weights. We challenge and relax a prevalent, yet often unrealistic, assumption in prior PAC-Bayesian frameworks: the statistical independence of weights. From this relaxation, we derive an improved PAC-Bayesian robust generalization bound. Our theoretical developments suggest that optimizing the second-order statistics of weights can substantially tighten this bound. We complement this theoretical insight by conducting an extensive set of experiments that demonstrate that S$^2$O not only enhances the robustness and generalization of neural networks when used in isolation, but also seamlessly augments other state-of-the-art adversarial training techniques. The code is available at https://github.com/Alexkael/S2O.

Annotating medical data for training AI models is often costly and limited due to the shortage of specialists with relevant clinical expertise. This challenge is further compounded by privacy and ethical concerns associated with sensitive patient information. As a result, well-trained medical segmentation models on private datasets constitute valuable intellectual property requiring robust protection mechanisms. Existing model protection techniques primarily focus on classification and generative tasks, while segmentation models-crucial to medical image analysis-remain largely underexplored. In this paper, we propose a novel, stealthy, and harmless method, StealthMark, for verifying the ownership of medical segmentation models under black-box conditions. Our approach subtly modulates model uncertainty without altering the final segmentation outputs, thereby preserving the model's performance. To enable ownership verification, we incorporate model-agnostic explanation methods, e.g. LIME, to extract feature attributions from the model outputs. Under specific triggering conditions, these explanations reveal a distinct and verifiable watermark. We further design the watermark as a QR code to facilitate robust and recognizable ownership claims. We conducted extensive experiments across four medical imaging datasets and five mainstream segmentation models. The results demonstrate the effectiveness, stealthiness, and harmlessness of our method on the original model's segmentation performance. For example, when applied to the SAM model, StealthMark consistently achieved ASR above 95% across various datasets while maintaining less than a 1% drop in Dice and AUC scores, significantly outperforming backdoor-based watermarking methods and highlighting its strong potential for practical deployment. Our implementation code is made available at: https://github.com/Qinkaiyu/StealthMark.

Image transmission and processing systems in resource-critical applications face significant challenges from adversarial perturbations that compromise mission-specific object classification. Current robustness testing methods require excessive computational resources through exhaustive frame-by-frame processing and full-image perturbations, proving impractical for large-scale deployments where massive image streams demand immediate processing. This paper presents DDSA (Dual-Domain Strategic Attack), a resource-efficient adversarial robustness testing framework that optimizes testing through temporal selectivity and spatial precision. We introduce a scenario-aware trigger function that identifies critical frames requiring robustness evaluation based on class priority and model uncertainty, and employ explainable AI techniques to locate influential pixel regions for targeted perturbation. Our dual-domain approach achieves substantial temporal-spatial resource conservation while maintaining attack effectiveness. The framework enables practical deployment of comprehensive adversarial robustness testing in resource-constrained real-time applications where computational efficiency directly impacts mission success.

As large language models (LLMs) transition to autonomous agents synthesizing real-time information, their reasoning capabilities introduce an unexpected attack surface. This paper introduces a novel threat where colluding agents steer victim beliefs using only truthful evidence fragments distributed through public channels, without relying on covert communications, backdoors, or falsified documents. By exploiting LLMs' overthinking tendency, we formalize the first cognitive collusion attack and propose Generative Montage: a Writer-Editor-Director framework that constructs deceptive narratives through adversarial debate and coordinated posting of evidence fragments, causing victims to internalize and propagate fabricated conclusions. To study this risk, we develop CoPHEME, a dataset derived from real-world rumor events, and simulate attacks across diverse LLM families. Our results show pervasive vulnerability across 14 LLM families: attack success rates reach 74.4% for proprietary models and 70.6% for open-weights models. Counterintuitively, stronger reasoning capabilities increase susceptibility, with reasoning-specialized models showing higher attack success than base models or prompts. Furthermore, these false beliefs then cascade to downstream judges, achieving over 60% deception rates, highlighting a socio-technical vulnerability in how LLM-based agents interact with dynamic information environments. Our implementation and data are available at: https://github.com/CharlesJW222/Lying_with_Truth/tree/main.

As generative models become increasingly diverse and powerful, cross-generator detection has emerged as a new challenge. Existing detection methods often memorize artifacts of specific generative models rather than learning transferable cues, leading to substantial failures on unseen generators. Surprisingly, this work finds that frozen visual foundation models, especially DINOv3, already exhibit strong cross-generator detection capability without any fine-tuning. Through systematic studies on frequency, spatial, and token perspectives, we observe that DINOv3 tends to rely on global, low-frequency structures as weak but transferable authenticity cues instead of high-frequency, generator-specific artifacts. Motivated by this insight, we introduce a simple, training-free token-ranking strategy followed by a lightweight linear probe to select a small subset of authenticity-relevant tokens. This token subset consistently improves detection accuracy across all evaluated datasets. Our study provides empirical evidence and a feasible hypothesis for understanding why foundation models generalize across diverse generators, offering a universal, efficient, and interpretable baseline for image forgery detection.

Personalized AI applications such as DreamBooth enable the generation of customized content from user images, but also raise significant privacy concerns, particularly the risk of facial identity leakage. Recent defense mechanisms like Anti-DreamBooth attempt to mitigate this risk by injecting adversarial perturbations into user photos to prevent successful personalization. However, we identify two critical yet overlooked limitations of these methods. First, the adversarial examples often exhibit perceptible artifacts such as conspicuous patterns or stripes, making them easily detectable as manipulated content. Second, the perturbations are highly fragile, as even a simple, non-learned filter can effectively remove them, thereby restoring the model's ability to memorize and reproduce user identity. To investigate this vulnerability, we propose a novel evaluation framework, AntiDB_Purify, to systematically evaluate existing defenses under realistic purification threats, including both traditional image filters and adversarial purification. Results reveal that none of the current methods maintains their protective effectiveness under such threats. These findings highlight that current defenses offer a false sense of security and underscore the urgent need for more imperceptible and robust protections to safeguard user identity in personalized generation.

Deep learning models are notoriously vulnerable to imperceptible perturbations. Most existing research centers on adversarial robustness (AR), which evaluates models under worst-case scenarios by examining the existence of deterministic adversarial examples (AEs). In contrast, probabilistic robustness (PR) adopts a statistical perspective, measuring the probability that predictions remain correct under stochastic perturbations. While PR is widely regarded as a practical complement to AR, dedicated training methods for improving PR are still relatively underexplored, albeit with emerging progress. Among the few PR-targeted training methods, we identify three limitations: i non-comparable evaluation protocols; ii limited comparisons to strong AT baselines despite anecdotal PR gains from AT; and iii no unified framework to compare the generalization of these methods. Thus, we introduce PRBench, the first benchmark dedicated to evaluating improvements in PR achieved by different robustness training methods. PRBench empirically compares most common AT and PR-targeted training methods using a comprehensive set of metrics, including clean accuracy, PR and AR performance, training efficiency, and generalization error (GE). We also provide theoretical analysis on the GE of PR performance across different training methods. Main findings revealed by PRBench include: AT methods are more versatile than PR-targeted training methods in terms of improving both AR and PR performance across diverse hyperparameter settings, while PR-targeted training methods consistently yield lower GE and higher clean accuracy. A leaderboard comprising 222 trained models across 7 datasets and 10 model architectures is publicly available at https://tmpspace.github.io/PRBenchLeaderboard/.

Graph neural networks (GNNs) are the predominant architecture for learning over graphs. As with any machine learning model, and important issue is the detection of adversarial attacks, where an adversary can change the output with a small perturbation of the input. Techniques for solving the adversarial robustness problem - determining whether such an attack exists - were originally developed for image classification, but there are variants for many other machine learning architectures. In the case of graph learning, the attack model usually considers changes to the graph structure in addition to or instead of the numerical features of the input, and the state of the art techniques in the area proceed via reduction to constraint solving, working on top of powerful solvers, e.g. for mixed integer programming. We show that it is possible to improve on the state of the art in structural robustness by replacing the use of powerful solvers by calls to efficient partial solvers, which run in polynomial time but may be incomplete. We evaluate our tool RobLight on a diverse set of GNN variants and datasets.

Within the PAC-Bayesian framework, the Gibbs classifier (defined on a posterior $Q$) and the corresponding $Q$-weighted majority vote classifier are commonly used to analyze the generalization performance. However, there exists a notable lack in theoretical research exploring the certified robustness of majority vote classifier and its interplay with generalization. In this study, we develop a generalization error bound that possesses a certified robust radius for the smoothed majority vote classifier (i.e., the $Q$-weighted majority vote classifier with smoothed inputs); In other words, the generalization bound holds under any data perturbation within the certified robust radius. As a byproduct, we find that the underpinnings of both the generalization bound and the certified robust radius draw, in part, upon weight spectral norm, which thereby inspires the adoption of spectral regularization in smooth training to boost certified robustness. Utilizing the dimension-independent property of spherical Gaussian inputs in smooth training, we propose a novel and inexpensive spectral regularizer to enhance the smoothed majority vote classifier. In addition to the theoretical contribution, a set of empirical results is provided to substantiate the effectiveness of our proposed method.

Vision Language Models (VLMs) have demonstrated impressive capabilities in integrating visual and textual information for understanding and reasoning, but remain highly vulnerable to adversarial attacks. While activation steering has emerged as a promising defence, existing approaches often rely on task-specific contrastive prompts to extract harmful directions, which exhibit suboptimal performance and can degrade visual grounding performance. To address these limitations, we propose \textit{Sequence-Level Preference Optimization} for VLM (\textit{SPO-VLM}), a novel two-stage defense framework that combines activation-level intervention with policy-level optimization to enhance model robustness. In \textit{Stage I}, we compute adaptive layer-specific steering vectors from diverse data sources, enabling generalized suppression of harmful behaviors during inference. In \textit{Stage II}, we refine these steering vectors through a sequence-level preference optimization process. This stage integrates automated toxicity assessment, as well as visual-consistency rewards based on caption-image alignment, to achieve safe and semantically grounded text generation. The two-stage structure of SPO-VLM balances efficiency and effectiveness by combining a lightweight mitigation foundation in Stage I with deeper policy refinement in Stage II. Extensive experiments shown SPO-VLM enhances safety against attacks via activation steering and preference optimization, while maintaining strong performance on benign tasks without compromising visual understanding capabilities. We will release our code, model weights, and evaluation toolkit to support reproducibility and future research. \textcolor{red}{Warning: This paper may contain examples of offensive or harmful text and images.}

Recent advances in Chain-of-Thought (CoT) prompting have substantially enhanced the reasoning capabilities of large language models (LLMs), enabling sophisticated problem-solving through explicit multi-step reasoning traces. However, these enhanced reasoning processes introduce novel attack surfaces, particularly vulnerabilities to computational inefficiency through unnecessarily verbose reasoning chains that consume excessive resources without corresponding performance gains. Prior overthinking attacks typically require restrictive conditions including access to external knowledge sources for data poisoning, reliance on retrievable poisoned content, and structurally obvious templates that limit practical applicability in real-world scenarios. To address these limitations, we propose POT (Prompt-Only OverThinking), a novel black-box attack framework that employs LLM-based iterative optimization to generate covert and semantically natural adversarial prompts, eliminating dependence on external data access and model retrieval. Extensive experiments across diverse model architectures and datasets demonstrate that POT achieves superior performance compared to other methods.

The rapid advancement of AI-generation models has enabled the creation of hyperrealistic imagery, posing ethical risks through widespread misinformation. Current deepfake detection methods, categorized as face specific detectors or general AI-generated detectors, lack transparency by framing detection as a classification task without explaining decisions. While several LLM-based approaches offer explainability, they suffer from coarse-grained analyses and dependency on labor-intensive annotations. This paper introduces RAIDX (Retrieval-Augmented Image Deepfake Detection and Explainability), a novel deepfake detection framework integrating Retrieval-Augmented Generation (RAG) and Group Relative Policy Optimization (GRPO) to enhance detection accuracy and decision explainability. Specifically, RAIDX leverages RAG to incorporate external knowledge for improved detection accuracy and employs GRPO to autonomously generate fine-grained textual explanations and saliency maps, eliminating the need for extensive manual annotations. Experiments on multiple benchmarks demonstrate RAIDX's effectiveness in identifying real or fake, and providing interpretable rationales in both textual descriptions and saliency maps, achieving state-of-the-art detection performance while advancing transparency in deepfake identification. RAIDX represents the first unified framework to synergize RAG and GRPO, addressing critical gaps in accuracy and explainability. Our code and models will be publicly available.