Large Audio-language Models (LAMs) have recently enabled powerful speech-based interactions by coupling audio encoders with Large Language Models (LLMs). However, the security of LAMs under adversarial attacks remains underexplored, especially through audio jailbreaks that craft malicious audio prompts to bypass alignment. Existing efforts primarily rely on converting text-based attacks into speech or applying shallow signal-level perturbations, overlooking the impact of human speech’s expressive variations on LAM alignment robustness. To address this gap, we propose StyleBreak, a novel style-aware audio jailbreak framework that systematically investigates how diverse human speech attributes affect LAM alignment robustness. Specifically, StyleBreak employs a two-stage style-aware transformation pipeline that perturbs both textual content and audio to control linguistic, paralinguistic, and extralinguistic attributes. Furthermore, we develop a query-adaptive policy network that automatically searches for adversarial styles to enhance the efficiency of LAM jailbreak exploration. Extensive evaluations demonstrate that LAMs exhibit critical vulnerabilities when exposed to diverse human speech attributes. Moreover, StyleBreak achieves substantial improvements in attack effectiveness and efficiency across multiple attack paradigms, highlighting the urgent need for more robust alignment in LAMs.

Large Language Models (LLMs) aligned with human feedback have recently garnered significant attention. However, it remains vulnerable to jailbreak attacks, where adversaries manipulate prompts to induce harmful outputs. Exploring jailbreak attacks enables us to investigate the vulnerabilities of LLMs and further guides us in enhancing their security. Unfortunately, existing techniques mainly rely on handcrafted templates or generated-based optimization, posing challenges in scalability, efficiency and universality. To address these issues, we present JailPO, a novel black-box jailbreak framework to examine LLM alignment. For scalability and universality, JailPO meticulously trains attack models to automatically generate covert jailbreak prompts. Furthermore, we introduce a preference optimization-based attack method to enhance the jailbreak effectiveness, thereby improving efficiency. To analyze model vulnerabilities, we provide three flexible jailbreak patterns. Extensive experiments demonstrate that JailPO not only automates the attack process while maintaining effectiveness but also exhibits superior performance in efficiency, universality, and robustness against defenses compared to baselines. Additionally, our analysis of the three JailPO patterns reveals that attacks based on complex templates exhibit higher attack strength, whereas covert question transformations elicit riskier responses and are more likely to bypass defense mechanisms.

Incorporating privacy regulations and business requirements, enterprises should securely manage unstructured textual data from diverse domains. Sensitive information classification is a critical component of data security, but it poses challenges due to complex textual contexts. With the ever-increasing data volumes, rule-based and lexicon-based classification methods require substantial maintenance and become inefficient. Additionally, supervised named entity recognition is not feasible for multi-domain sensitive information classification due to numerous type-specific annotations and the scarcity of training data. This work presents ToSIC, a new sensitive information classification approach to tackle these problems. Specifically, ToSIC simplifies multi-domain type labeling through a hierarchical structure and employs a rapid learning mechanism based on prototype distance classification to mitigate extensive training data. Experiments indicate that ToSIC outperforms existing methods by significantly reducing inference time and error rates while ensuring classification performance. Moreover, ToSIC demonstrates exceptional expandability to unknown types, showcasing potential value in practical applications.

Content Delivery Networks (CDNs) rely heavily on caching algorithms to minimize content delivery latency and optimize network performance. While machine learning approaches have emerged as promising solutions for handling complex request patterns in caching systems, current learningbased caching methods face critical limitations in processing granularity and operational efficiency. Existing approaches either process requests in coarse-grained time windows or struggle with throughput bottlenecks during object-level operations. To address these challenges, we present UniCache, a novel batchlevel content caching algorithm that balances processing granularity and system efficiency. UniCache introduces a Batch Queue architecture coupled with specialized Batch-level Inference components, enabling high-throughput processing while providing fine-grained request information. This design prevents both suboptimal caching decisions and request accumulation delays during prediction phases. Furthermore, UniCache overcomes the common limitation of treating admission and eviction policies as separate entities, by implementing a unified model that jointly optimizes both processes based on object popularity patterns. The integration of tiered cache storage enhances the system's resilience to prediction inaccuracies while facilitating effective identification and retention of popular objects. Experimental evaluation on Wiki CDN and Tencent Photo datasets demonstrates that UniCache achieves$\mathbf{1 2 \%} \sim \mathbf{5 3 \%}$improvement in Object Hit Ratio (OHR) compared to state-of-the-art methods while maintaining real-time processing capabilities. Comprehensive ablation studies validate the effectiveness of each architectural component in the overall system design.

Graph Neural Networks (GNNs) have achieved significant success in anomaly detection across various domains. However, existing GNN approaches face notable challenges when applied to directed multigraphs where edge attributes serve as the sole source of information. This limitation complicates the learning of meaningful node representations and the accurate detection of anomalies, particularly in transaction networks. Moreover, existing methods typically suffer from dimensionality explosion and suboptimal use of structural and semantic information, leading to reduced detection performance. To address these challenges, we propose Feature Reconstruction for Anomaly Detection on Directed Multigraphs (FRAD-DM), a novel preprocessing framework tailored to enhance graph representations in node feature-absent scenarios. FRAD-DM employs advanced edge feature derivation techniques, temporal and structural subgraph analysis for node representation generation, and a reinforcement learning-based adaptive feature selection mechanism. This comprehensive framework optimizes feature spaces by balancing informativeness, task relevance, and computational efficiency, ensuring robust and scalable anomaly detection. Extensive experiments on real-world datasets under barely supervised settings demonstrate the effectiveness of FRAD-DM. By integrating it into various GNN architectures for node-level and edge-level anomaly detection tasks, negative-class F1-scores improve by 5.44% to 20.98%. These results highlight its capability to address the challenges of directed multigraphs in practical applications.

Crop rotation is subsidized by the government because of its many advantages. Monitoring whether crop rotation is beneficial for agricultural management, and can also provide a reference for government subsidy policies for crop rotation. In this paper, we propose an unsupervised object-oriented crop rotation detection method using time-series polarimetric SAR (PolSAR) data. On the one hand, we construct the change detection matrix based on the likelihood ratio test (LRT) distance to perform temporal filtering. Then, the pixel-level temporal change image is generated using Shannon entropy and maximum between-class variance (OTSU). On the other hand, we perform temporal segmentation on time-series PolSAR images to obtain superpixel results. Finally, the object-level crop rotation results are obtained with the probabilistic label relaxation (PLR) model. 42 Sentinel-1 dual-polarization SAR datasets during 2018 and 2019 are selected for detecting crop rotation changes on farms within Jinchang, China. Experimental results show that the crop rotation detection accuracy andKappacoefficients of this method can reach 96.21% and 0.8989, respectively.

The superpixel generation is a key step for object-based classification and change detection. For the time-series polarimetric synthetic aperture radar (PolSAR) superpixel generation, the traditional polarimetric similarity measure based on the joint covariance matrix has limitations in discriminating different time-series similarity sequences with different fluctuations. Besides, in the traditional time-series PolSAR superpixel generation methods, it is difficult to determine the tradeoff factor between polarimetric and spatial similarity. In this article, an adaptive time-series PolSAR superpixel generation method based on the simple local iterative clustering (SLIC) is proposed, named time-series polarimetric SAR (TSPol)-adaptive simple local iterative clustering (ASLIC). There are three main improvements. First, a novel time-series polarimetric similarity measure based on the root mean square (rms) is proposed. Multitemporal polarimetric statistical information is combined to describe the polarimetric proximity between pixels, referring to the rms of the multitemporal proximities. Second, an edge detection method based on the stacked 2-D Gaussian-shaped (s2-D GS) window is proposed to initialize the central seeds for superpixel generation. Third, an improved SLIC clustering similarity combined with the time-series polarimetric, time-series power, and spatial similarities is proposed. Meanwhile, a homogeneity factor is applied to adaptively balance the relative weights of various similarities. We use eight Radarsat-2 quad-polarization synthetic aperture radar (SAR) images and 14 Sentinel-1 dual-polarization SAR images to evaluate the effectiveness. The results show our similarity measure and superpixel generation results are superior to those of the traditional methods. For example, as for the Radarsat-2 data, the improvement of the boundary recall by the proposed similarity measure and homogeneity factor is about 4% and 10%, respectively.