Large Language Models (LLMs) have achieved remarkable performance across a wide range of mathematical benchmarks.However, concerns remain as to whether these successes reflect genuine reasoning or superficial pattern recognition.Existing evaluation methods, which typically focus either on the final answer or on the intermediate reasoning steps, reduce mathematical reasoning to a shallow input-output mapping, overlooking its inherently multi-stage and multi-dimensional cognitive nature.Inspired by Pólya's problem-solving theory, we propose SMART, a benchmark that decomposes mathematical problem-solving into four cognitive dimensions: Semantic Understanding, Mathematical Reasoning, Arithmetic Computation, and Reflection & Refinement, and introduces dimension-specific tasks to measure the corresponding cognitive processes of LLMs.We apply SMART to 22 state-of-the-art open-and closed-source LLMs and uncover substantial discrepancies in their capabilities across dimensions.Our findings reveal genuine weaknesses in current models and motivate a new metric, the All-Pass Score, designed to better capture true problem-solving capability.

Advanced Persistent Threats (APTs) have become increasingly sophisticated and covert, necessitating the acquisition of an overall view of the rapidly evolving cyber threat landscape by security defenders. However, integrating threat intelligence from diverse sources poses significant challenges due to limited labeled data and noise interference. To address the requirement for the early detection of APT attacks, this paper introduces a lightweight framework named Warning-Graph, based on threat intelligence modeling. Warning-Graph leverages a limited set of IoCs to infer the type of ongoing APT attack. Initially, attack-related infrastructure nodes are modeled as a heterogeneous information network. Subsequently, heterogeneous graph contrastive learning is employed for pre-training. Two asymmetric graph encoders are constructed to obtain node embeddings without the need to generate negative samples or labeled data. In addition, a loss function based on the information bottleneck is specifically employed to reduce the noise in the original graph. In downstream tasks, multiclass classifiers are trained using embedding representations with fewer labeled samples. Experimental results demonstrate that the proposed framework achieves a 3- to 5-point increase in identification performance for APT attack types compared to baselines, while utilizing fewer labeled samples.

Recently, contrastive learning has emerged as a promising approach for multi-view clustering (MVC), as it enforces cross-view consistency and leverages complementary information from different views to enhance the analysis of heterogeneous data. However, traditional contrastive MVC methods suffer from an inherent limitation: their one-to-many contrast mechanism induces the False Negative Problem (FNP), where semantically similar intra-class instances are erroneously repelled. This phenomenon compromises intra-class consistency and ultimately degrades clustering performance. To overcome this issue, we propose a novel Pseudo lA bel gU ided univerS um lE arning (PAUSE) framework for robust multi-view clustering. Specifically, PAUSE operates in two synergistic stages: (1) A warm-up stage that employs dual contrastive learning to generate reliable pseudo-labels, establishing robust semantic relationships; (2) A fine-tuning stage that synthesizes universum samples via Mixup between anchor instances and out-of-class centroids, guided by the acquired pseudo-labels. This unique mechanism constructs generalized negative classes that expand inter-class margins while preserving intra-class cohesion. Crucially, the widened decision boundaries prevent misclassification of displaced intra-class instances, effectively circumventing FNP without requiring explicit negative pair correction. We further devise a robust universum contrastive loss that explicitly enforces cross-view consistency through adaptive boundary constraints. Extensive experiments on five multi-view benchmarks demonstrate that our PAUSE consistently outperforms 11 state-of-the-art multi-view learning methods. Our code is accessible at: https://github.com/xixi-555/PAUSE_main_code.

Tor is widely used to facilitate anonymous web communication, but it remains vulnerable to Website Fingerprinting (WF) attacks. Although deep learning-based WF attacks have shown promising results, they typically rely on large-scale labeled data and assume consistent conditions between training and deployment. These assumptions limit their practical applicability in real-world scenarios, where data scarcity and domain shifts are common. To address these challenges, recent research has focused on Cross-Domain Few-Shot Website Fingerprinting (CDFSWF), a more realistic yet challenging setting. Existing efforts mainly leverage data augmentation or feature alignment techniques. While data augmentation can mitigate sample scarcity, it often fails to capture true distributional variability. In contrast, many feature alignment WF methods overlook the semantic structure of class relationships, reducing their effectiveness in the target domain. In this paper, we propose CoDA, a novel method designed to improve cross-domain robustness in CDFSWF. CoDA integrates supervised contrastive pre-training, hierarchical flow attention, and prototype-based classification to effectively model semantic traffic structures under domain shifts. Furthermore, a Dual Confidence Alignment (DCA) strategy is introduced during fine-tuning to adaptively align semantic structures. Extensive experiments across various cross-domain scenarios show that CoDA consistently outperforms state-of-the-art baselines in both closed-world and open-world settings.

The one-time pad secure transmission based on wireless channel reciprocity (CR-OTP) has drawn great attention recently due to its capability of providing perfect secrecy of data, as well as the modulation information. However, existing CR-OTP schemes encounter both reliability and security challenges as their assumptions of channel reciprocity and randomness are not always well satisfied in practical application scenarios. To tackle these issues, we propose a dynamic constellation obfuscation (DCO) method that obfuscates the plaintext by rotating its constellation dynamically. This kind of analog encryption method is proven to be more robust than the existing digital exclusive OR (XOR) encryption method as the former achieves a lower symbol error rate (SER) by reducing the double quantization loss to one. The rotation pattern is jointly dependent on the channel state information (CSI) and the previous message, which guarantees the randomness of the rotation pattern subjected to environmental drifts. Only the legitimate receiver that correctly recovers the previous message correctly and observes a similar CSI is able to decode the newly transmitted message. We proved that the secrecy capacity of the proposed DCO method is higher than that of the state-of-the-art. Simulation results confirm that the proposed method delivers superior performance regarding secrecy capacity and SER, achieving a 4.5 dB signal-to-noise ratio (SNR) gain at a SER of 0.1; moreover, when the secrecy capacity is 0.1, the main channel SNR gain reaches 6.5 dB when the wiretap channel SNR is 20 dB.

Hidden services (HS) allow users to experience anonymity, but they also provide shelter for criminal activities. The widespread attention towards deanonymizing HS has brought website fingerprinting attack (WFA) into the spotlight, which is considered highly promising. However, most HS websites are designed simply and have high similarity in resource structures, making it difficult to represent the HS access traffic well, and existing work often directly applies traffic representation methods in the field of web research, resulting in poor effects of the model. Besides, features of HS access traffic are closely related to the resource access sequence of websites. Current studies build models based on convolutional neural network (CNN), ignoring the global correlation of HS access traffic parts. To address the short-comings, we have proposed an efficient WFA to deanonymize HS, and named it NuanceTracker. The contribution of our work lies in three points. Firstly, a burst-based HS fingerprint generation algorithm is proposed to describe the sequence of HS access traffic. Secondly, we propose NuanceTracker, which is designed by introducing multi-scale global attention (MGA) into a basic CNN model for global information extraction. Finally, comparison experiments are conducted in closed-world and open-world scenarios. Our NuanceTracker has proven to outperform three state-of-the-art WFA methods.

Tor has provided hidden services (HS) and protected the anonymity of the Web server with hidden service directory servers. Some criminals use hidden services to engage in illegal activities, such as anonymous transactions, pirated distribution, hacking, etc. In order to protect the security of cyberspace, hidden service traffic needs to be deanonymized. Artificial intelligence-based methods have become the most promising, but there are still two shortcomings in current research work. First, some of them mainly uses the size and direction sequence of the data packet as the input to complete the recognition, without mining the features of network traffic from many views. Second, they extract information from different view, but just concatenate them together instead of fuse them densely. Therefore, in this paper we propose a Tor hidden service traffic identification method with multi views named TriViewNet. TriViewNet extracts information from three different views, local flow, TLS layer, and TCP layer for identification ad fuses them with Tri-attention module. By comparing with state-of-the-art models, the results show that our TriViewNet outperforms in the recognition of Tor HS traffic.

Physical layer encryption (PLE) has emerged as a promising technique to secure wireless communications. Different from conventional cryptography implemented at higher layers, PLE exploits the randomness of wireless channels to adjust symbol patterns at the physical layer, by which both data and modulation information can be protected. However, existing PLE schemes face challenges of security and robustness in practical usage. In a slowly varying environment, the constellation variation is negligible, which results in the vulnerability of PLE to the differential attack. Moreover, the decryption error rate of PLE is high when the channel reciprocity is not ideal. To tackle these problems, we exploit data randomness to enhance the dynamics of constellation variations between adjacent frames. Then we utilize analog-based encryption instead of digital-based encryption to dynamically rotate constellation, which reduces quantization loss and improves robustness to channel phase errors. Simulation results verify that the proposed scheme can effectively resist the differential attack and provide approximately a 4.5 dB gain when the bit error ratio (BER) is 0.001.