Reconstructing past events in IT systems is a critical bottleneck in forensic investigations, consuming valuable time from investigators. It requires meticulous analysis of complex digital traces in an environment where attackers may try to erase traces. For example, deletion of digital artifacts is an anti-forensic technique used to jeopardize the success of forensic investigations.To address these challenges, we introduce Investigator Copilot, a novel framework that automates post-mortem event reconstruction using explainable machine learning. To overcome the general scarcity of datasets, Investigator Copilot replays realistic events on virtual machines, and creates datasets by extracting, normalizing and labeling traces from corresponding hard disks. Using these datasets, Investigator Copilot trains human-interpretable decision tree stumps that evaluate digital evidence and combines these binary classifiers in Forensic Forests. Forensic Forests utilize an adjusted voting scheme to provide robust event reconstruction even when faced with deleted evidence.We evaluate our approach by executing 2100 events on 50 virtual machines, training Forensic Forests and measuring their event reconstruction performance on previously unseen data. Our results demonstrate that tree-based classifiers perform exceedingly well in event reconstruction. When measuring reconstruction performance on manipulated evidence, we observe that Forensic Forests significantly outperform the state-of-the-art, which positions them as a valuable tool for investigators. Our findings indicate that automated frameworks such as Investigator Copilot can contribute to the efficiency and robustness of forensic analyses, and may save scarce resources of human investigators.
Fuzzy support vector machine (FSVM) effectively enhances the model’s robustness to common noise and outliers by introducing a sample membership mechanism. To further improve the model’s performance in complex scenarios such as data poisoning and adversarial noise, researchers have proposed the curriculum learning-based fuzzy support vector machine (CLFSVM), which optimizes the training process through a curriculum learning strategy. However, the performance of this model heavily depends on the quality of the initial core sample set, and an improper initial selection can lead to a significant degradation in model performance. To address this issue, we propose an innovative framework, the granular ball and curriculum learning-based FSVM (GB-CLFSVM), which includes three key components. First, it constructs a multi-stage training paradigm from easy to hard based on curriculum learning theory, ensuring the model gradually adapts to complex data distributions. Second, it employs granular balls to automatically select a set of core samples, building a small and high-quality training set and generating a robust standard SVM as the initialization baseline for FSVM. Finally, it designs an adaptive composite weighting mechanism to dynamically calculate sample fuzzy memberships. By comprehensively considering local sample density and class information, this mechanism assigns fine-grained weights to samples of varying complexity, thereby optimizing the iterative process of curriculum learning. Experimental results show that GB-CLFSVM significantly outperforms the CLFSVM baseline model in both classification accuracy and noise resistance, demonstrating exceptional generalization ability, especially in data scenarios with noisy features and labels.
Black-box adversarial attacks bypass internal target model information, broadening their applicability and making them a critical area in deep learning security. However, most existing methods prioritize attack success over perturbation imperceptibility, resulting in visually detectable perturbations. Addressing this, we propose the Grad-CAM-guided Perceptibility-Reduced Ensemble Attack (GPEA), integrating ensemble learning and Gradient-weighted Class Activation Mapping (Grad-CAM) to improve both attack success and perturbation imperceptibility. GPEA comprises two modules: Dynamic Weight Ensemble (DWE) and Grad-CAM-Based Perturbation Restriction (GPR). DWE introduces a novel weight initialization and adjustment scheme, optimizing surrogate model contributions to enhance black-box attack success. GPR applies Grad-CAM to restrict perturbations to critical image areas, markedly reducing perceptibility. Experimental results demonstrate the effectiveness of GPEA in balancing attack success and imperceptibility across various datasets and models. On VOC dataset, GPEA improves the black-box attack success rate by up to 8.4% over state-of-the-art methods; across five target models, it achieves average improvements of 18.5%, 16%, 36.4%, 33.9%, and 26.8% in SSIM, PSNR, LPIPS, l(0), and l(2) metrics, respectively.
Containers have become a widely used mechanism for isolation of software workloads on the same physical platform in cloud and edge computing scenarios. However, the isolation guarantees provided by containers result from multiple, interacting mechanisms for process isolation and virtualization at the kernel level. These mechanisms are often configured through a container engine with its own engine-level configuration semantics. Consequently, deciding whether a particular container configuration provides the level of isolation required in a given scenario is a non-trivial taskThe present work proposes an approach to assess the isolation characteristics of a container in an engine-agnostic way. To this end, we evaluate the container and host views of the configuration of process isolation mechanisms for a given container at the kernel user-space API-level. The approach for configuration assessment we propose makes the isolation characteristics that are provided by a container transparent. This enables, for instance, container engine administrators to assess container configurations prior to large-scale deployment in an efficient manner. Also, the proposed approach makes container isolation characteristics comparable among container engines and enables verification of the container isolation configuration at runtime.
Large vision-language models (LVLMs) have demonstrated remarkable capabilities across diverse multimodal tasks, yet remain vulnerable to adversarial attacks, posing significant challenges to their secure deployment. To reveal their vulnerabilities and promote robust multimodal learning, we propose TSRP, a targeted and transferable attack framework utilizing two-stage region-specific perturbations. In Stage One, TSRP optimizes perturbations in critical regions using Grad-CAM-guided contrastive learning on eight large CLIP variants, enhanced with extensive image data augmentation to improve robustness and transferability. In Stage Two, it refines perturbations in non-critical regions via the vision encoder of GLM-4V, ensuring efficient use of the image space while preserving perceptual similarity. Extensive experiments on multiple state-of-the-art LVLMs show that TSRP achieves an average attack success rate of 89.9% with the lowest LPIPS score of 0.1536, significantly outperforming existing methods. These results demonstrate that region-specific, multi-stage perturbation optimization substantially enhances both attack effectiveness and transferability, offering new insights for adversarial research on LVLMs.
Dynamic trust evaluation in online social networks is of great significance for applications such as recommendation systems, financial risk control, and e-commerce, but it remains a challenging research problem. Most existing methods are based on static network topologies or simple temporal modeling, making it difficult to effectively capture the dynamic evolution laws of trust relationships. To address this issue, this paper proposes a dynamic graph neural network-based trust evaluation model, TeGCN, for predicting dynamic trust relationships in social networks. TeGCN consists of a spatial aggregation unit, a gated temporal evolution unit, and a joint prediction unit. In the spatial aggregation unit, node spatial feature representations are generated through linear transformation after concatenating features of neighbor nodes and forward/backward edge attribute information via average aggregation. The gated temporal evolution unit employs a Gated Recurrent Unit (GRU) to model the temporal evolution of trust relationships and introduces a dynamic weight adjustment mechanism to enhance feature extraction at key time steps. Finally, the joint prediction unit integrates spatio-temporal features, performs nonlinear mapping through a multi-layer perceptron (MLP), and outputs the probability distribution of node trust levels to achieve end-to-end dynamic trust prediction. Experiments based on two real dynamic trust network datasets, Bitcoin-Alpha and Bitcoin-OTC, verify the effectiveness of TeGCN. The results show that TeGCN significantly outperforms existing baseline models in key metrics such as F1-score, AUC, and ACC, especially in highly dynamic scenarios.
With the advancement of deep learning, facial forgery techniques have achieved remarkable breakthroughs and demonstrate broad application potential in fields such as film production. However, their associated societal risks cannot be overlooked. Notably, recent studies indicate that these deepfake detection models may exhibit unfair performance disparities across different demographic groups (e.g., race, gender). It is therefore crucial to develop efficient, accurate, and fairness-aware deepfake detection methods. To address this challenge, we propose TriDNet, a novel deepfake detection algorithm that integrates 3D feature reconstruction with generalizable feature disentanglement. This approach effectively disentangles generalized features and improves detection accuracy. Specifically, our model first extracts 3D facial reconstruction features and incorporates them into the original images for data augmentation. We then construct a 3D feature disentanglement model consisting of three key components: a feature disentanglement module, a generalization loss module, and an optimization module, to perform deepfake detection. Experimental results demonstrate that our method achieves significant performance improvements in both in-domain and cross-domain datasets, while maintaining fairness across different demographic groups.
With the rapid advancement of cloud computing technologies, microservice architectures have emerged as the de facto standard for modern software systems. However, their highly dynamic and complex nature poses significant challenges for anomaly detection. Although recent diffusion-based models have demonstrated promising capabilities in modeling multi-variate time series, they often overlook the inherent multi-scale spatiotemporal correlations present in microservice systems. This oversight limits their capacity to fully capture system-wide behaviors, thereby constraining detection performance. To address this issue, we propose a novel framework for anomaly detection based on imputed diffusion models, termed Multi-Scale Graph Imputed Diffusion Model. Building upon the imputation-reconstruction paradigm of standard ImDiffusion, our approach introduces a multi-scale spatiotemporal correlation module that is seamlessly integrated into the denoising network. This module employs parallel multi-scale graph convolutions and temporal convolutions to explicitly capture inter-service spatial dependencies from the microservice dependency graph and multi-granular temporal dynamics from service metric time series. By conditioning the diffusion process on richer spatiotemporal contexts, MSG-ImDiffusion enhances the model's sensitivity to anomalies that deviate from expected temporal and spatial patterns. Extensive experiments conducted on the open-source HipsterShop microservice benchmark demonstrate that our proposed model significantly outperforms existing baselines, including ImDiffusion, in terms of both detection accuracy and F1-score. These results validate the effectiveness of incorporating multi-scale spatiotemporal information into diffusion-based anomaly detection frameworks for microservice systems.
Microservice architectures have gained widespread adoption in cloud-native environments due to their flexibility and scalability. However, these architectures pose significant challenges in the automated generation and dynamic updating of fine-grained security policies. This paper presents LLM2policy, a novel framework that utilizes large language models (LLMs) for end-to-end automated security policy generation. LLM2policy extracts microservice entity information from deployment YAML files and identifies RPC call relationships from distributed tracing data, consolidating this information into a structured knowledge base. This knowledge base is then used to automatically generate Istio-compatible access control policies in YAML format, enabling dynamic policy updates. Evaluation results from five benchmark microservice systems demonstrate that LLM2policy achieves 100% accuracy in entity recognition and dependency extraction, over 98.81% accuracy in semantic extraction by the LLM, and unit test pass rates ranging from 93.75% to 100% for the generated policies. Furthermore, attack simulations confirm that the generated policies effectively mitigate unauthorized access, highlighting the practical applicability and robustness of LLM2policy in automated cloud-native security policy management.
The inherent tension between transactional privacy and auditing compliance presents a significant challenge to the widespread adoption of permissioned blockchain systems, particularly in highly regulated sectors such as finance. Existing approaches often suffer from inadequate privacy protection, rigid audit frameworks, or excessive computational and user interaction overhead. This paper introduces a novel permissioned blockchain transaction system specifically designed to address these challenges. Our design guarantees complete anonymity for sender and receiver identities as well as transaction amounts. Additionally, we implement a streamlined, non-interactive audit mechanism that allows authorized entities to inspect transactions without participant involvement. A key innovation is certifier-aided balance reconciliation, which significantly enhances transaction efficiency, usability, and scalability by eliminating the need for direct communication between participants. Through a formal security proof within the universal composable framework, we demonstrate the robust ability of the system to reconcile privacy requirements with compliance obligations.
Federated learning (FL) has emerged as a privacy-preserving alternative to traditional machine learning, enabling collaborative model training without centralizing users’ data. While this decentralized approach enhances privacy, it introduces new trust assumptions. Specifically, that users honestly participate in the protocol. Simultaneously, the growing emphasis on fairness in machine learning has led to the development of fairness-aware FL schemes, which aim to ensure equitable model performance across diverse user groups. However, existing solutions for fair FL schemes rely on users truthfully reporting fairness-related metrics. This trust assumption opens the door to malicious behavior: users may manipulate these statistics to influence the global model, degrading its fairness, and/or its overall accuracy. Current frameworks lack mechanisms to detect or prevent such adversarial manipulation. This paper addresses this critical gap through two main contributions. First, as an example to illustrate our point, we empirically demonstrate the vulnerability of a representative fairness-aware FL framework to targeted attacks that exploit unverified fairness computations. Second, and most importantly, we propose a novel scheme (that can be integrated into existing fair FL schemes) that augments fairness-aware FL with verifiability. Our solution enables the detection of dishonest participants without compromising user privacy, thus strengthening the robustness of fairness-aware FL schemes.
Serverless computing has gained widespread popularity among developers due to its low cost, fine-grained deployment, and management-free operation. However, when deploying serverless applications in practice, a single function is often insufficient to fulfill complete application requirements. This has led to the emergence of serverless workflows, which orchestrate a series of related serverless functions according to predefined logic. Through our investigation of existing serverless workflow platforms, we identify two major security limitations. First, current serverless workflows cannot guarantee execution integrity—they are unable to detect changes in the function execution order and lack mechanisms to defend against workflow-targeted denial-of-service (DoS) attacks. Second, identity management is typically coarse-grained, often resulting in over-privileged access and lacking support for function-level access control.To address these issues, we propose and implement the LASEFlow, a label-aware security enhancement framework for serverless workflows. A sequential function execution chain is designed based on the Platform Configuration Register (PCR) technique to guarantee the integrity of the execution of workflow in LASEFlow. In addition, a fine-grained function-level access control mechanism is designed to prevent privilege abuse in work-flows. The evaluation demonstrates the effectiveness of LASEFlow against workflow attacks, with an overhead of less than 4% in performance.
Tor is one of the most widely adopted anonymity networks, yet its anonymity can be undermined by adversaries through flow correlation attacks. Current mainstream technologies focus on exploiting the sequence characteristics of packet lengths and timestamps to execute attacks. However, the padding mechanism of the Tor network and time delays caused by multi-hop relays obscure these single-modal features. Additionally, the diversity of network services and the randomness of user behavior result in sparse packet distributions, which impact model training and inference. In this paper, we propose SSRCorr, a novel self-supervised learning framework for flow correlation attacks, incorporating the Flow Feature Aggregation (FFA) module and Global-Local Fusion (GLoF) Encoder to address these challenges. Firstly, we construct a Byte-based Traffic Aggregation Matrix (BTAM) by integrating time and length sequences and applying two data augmentation methods tailored for Tor flow correlation, thereby reducing the impact of Tor network noise on attack effectiveness. Secondly, we employ GLoF to extract features from the output by FFA and fuse the global context information of the traffic, thus mitigating the impact of low-information traffic on model performance. Experiments show that SSRCorr achieves a TPR of 96%, surpassing other methods, and maintains robust performance under temporal drift and obfuscation, supporting future research on countering anonymity system defenses.
Wireless sensor networks (WSNs) are a crucial component of the Internet of Things (IoT). In recent years, the increasing number of smart devices communicating through WSNs has fueled growing research in this area. While WSNs hold great potential for development, they also face several challenges, such as ensuring key security and facilitating key updates. In this paper, we propose a threshold distributed key generation (TDKG) scheme based on verifiable secret sharing. Additionally, we present a dynamic key management scheme that utilizes the vector commitment to ensure that the secret key generated by smart devices cannot be arbitrarily altered. Subsequently, we describe the ECC-based instantiation schemes for TDKG and key management. Besides, the security analysis demonstrates that our proposed scheme offers enhanced security properties compared to existing approaches, while also achieving good performance in terms of time and communication costs.
Early classification of encrypted traffic is crucial for time-sensitive tasks such as intrusion detection and cyberspace situation awareness. Dissimilar flows can be distinguished with a small number of packets, while similar flows require more packets. The existing early classification methods use the same number of packets to classify all flows, which leads to many flows cannot be classified early. In this paper, we propose a method to classify TLS flows by concatenation of multiple models. For flows that fail to be classified, a larger number of packets are used to classify them in the subsequent models in turn until they are successfully classified. Compared with existing methods that rely on unstable features such as time series, the proposed method can accurately classify flows by features that are easy to extract and maintain stable values. Experiments show that the proposed framework has the highest classification efficiency and accuracy compared with three baseline methods, including a state-of-the-art early classification method.
Protocols serve as the foundation for communication between devices and systems in modern computing environments. However, their widespread use also exposes them to various security threats. A flaw in protocol implementation can be exploited to compromise system security, leading to severe disruptions or unauthorized access. Fuzzing has become a widely used technique for protocol vulnerability detection. However, the existing protocol fuzzing approaches often lack effective guidance strategies for exploring the complex protocol state space, which leads to limited coverage and suboptimal testing efficiency.In this paper, we propose BayesFuzz, a Bayesian-guided greybox fuzzing approach for stateful protocols. It maintains a probability table to record the probability of sending different messages and triggering different state transitions in the current state. This table is updated via Bayesian inference based on feedback from each fuzzing iteration. Accordingly, BayesFuzz is able to select the optimal message for the current state during test case generation, thereby improving fuzzing efficiency in complex stateful protocols. Experimental results confirm the effectiveness of BayesFuzz. Compared with state-of-the-art fuzzers BooFuzz and AFLNET, BayesFuzz increases branch coverage by averagely 23.18% and 45.3% within 24 hours. Furthermore, it successfully discovered an unknown vulnerability in the MQTT protocol implementation.
As Large Language Models (LLMs) are integrated into high-stakes societal functions, their inherent biases have evolved from ethical concerns into critical, exploitable security vulnerabilities that undermine system integrity and trust. Traditional safety evaluations often fail to detect these subtle, context-dependent flaws. To address this, we introduce a scalable red-teaming framework designed to systematically attack and expose latent gender bias vulnerabilities in foundational models. Our framework operationalizes bias as an exploit, leveraging three distinct attack patterns—Latent Bias Elicitation, Forced-Choice Discrimination, and Stereotype-Amplifying Narrative Generation—to bypass safeguards and compel biased outcomes. We deployed this framework in a large-scale offensive campaign against a cohort of globally significant models, including the GPT, Claude, Gemini, and leading Chinese foundational model series. The attacks successfully manipulated all targets into producing statistically significant discriminatory outputs, proving that inherent bias is an operationally exploitable vulnerability. We discovered asymmetric weaknesses: English-centric models were attacked to exhibit strong male bias in Chinese contexts, while Chinese-centric models were vulnerable to similar male-biased exploits across both languages. This work provides concrete demonstration of socio-cultural bias as a potent and scalable attack vector, establishing the necessity of adversarial red-teaming for building trustworthy AI. All attack data and scripts are open-sourced to facilitate further security audits.
Passwordless authentication is rapidly reshaping digital identity management, with passkeys emerging as a key technology prepared to replace traditional passwords. Built on the FIDO2 and WebAuthn standards, passkeys promise secure, phishing-resistant, and user-friendly authentication. Despite increasing industry adoption, their real-world interoperability across browsers and devices remains untested. This paper presents one of the first empirical studies evaluating passkey implementations in Chrome, Safari, Edge, and Firefox across four passkey types: iCloud Keychain, browser profile credentials, phone or tablet-based QR code workflows, and USB security keys. We propose a structured evaluation framework and introduce metrics to assess cross-browser credential reuse, cross-device synchronization, and roaming authenticator interoperability. Our results expose cross-browser inconsistencies driven by non-uniform credential management, consent handling, and error reporting. The study contributes a structured evaluation framework, reproducible criteria, and interoperability metrics that benchmark passkey readiness across major browsers.
