Trusted Execution Environments (TEEs) provide an efficient foundation for ciphertext computation through hardware-enforced isolated execution, enabling secure multi-party data processing. Integrating TEEs with cryptographic techniques allows sensitive data to be processed without exposing the plaintext, thereby supporting collaborative computation in untrusted environments. These lightweight and efficient security mechanisms have driven widespread interest and research into TEE-based ciphertext computation schemes. However, existing studies often overlook the impact of participating entity behaviors on overall security, lacking a systematic characterization of the complex threat landscape introduced by TEEs. To address this gap, we present a survey of TEE-assisted computation over ciphertext from this behavioral perspective, providing an in-depth analysis of the limitations of current trust models. To comprehensively characterize the complex threats introduced by TEEs, we propose a security framework driven by these entity behaviors. By defining the capabilities and goals of each entity, this framework systematizes the threat models of TEE-assisted ciphertext computation systems into three core trust assumptions: the honest, the semi-honest, and the malicious assumption. Based on this framework, we systematically classify TEE-assisted ciphertext computation protocols across typical scenarios. We evaluate their design features and security under different trust assumptions, and briefly discuss potential open challenges in this domain.
As the field of personal lending grows exponentially, the risk of credit default has become an issue of growing significance. Meanwhile, responsible and fair personal credit evaluation in safe environments is a key area of research in the field of financial technologies. Even the current strategies have a few problems in the real-world application: the imbalanced classes of credit data creates low detectability of delinquent cases, a centralized scoring solution does not provide credible verification methods, which makes them easily manipulated, and machine learning model decision-making is not typically transparent.In resolving these problems, this study introduces a verifiable and interpretable model of personal credit assessment that combines ensemble learning, blockchain-based data notarization, and zero-knowledge proofing. To begin with, a credit assessment model using soft-voting ensemble, with SMOTE oversampling is developed, to integrate the complementary advantages of random forest and gradient boosting classifier and thus would be useful in enhancing the detection of delinquent users. Second, a zero-knowledge proof system developed under the Groth16 protocol and integrated with blockchain notarization to make the verifiability and the immutability of the scoring outcomes accessible to protect data privacy. Third, the entire SHAP feature attribution vectors are hashed with the hash-digest algorithm SHA-256 and stored on-chain, providing end-to-end auditability over scoring results to the underlying rationale decision and raising decision transparency and regulatory acceptability.In the experiments conducted on UCI Germany credit dataset, the suggested method proves to be more efficient than the baseline models in major measurements including full term and Recall. The Recall was improved by approximately 30 percentage points and the AUC increased from 0.8938 to 0.9024. Both on-chain audit and zero-knowledge proof verification rates are 100%, and SHAP notarization completion is 100%, which show general benefits of the framework in predictive accuracy, privacy protection, model explainability, and result verifiability. The work presents a practicable technical route of the personal credit assessment frameworks within the financial regulation realms.
Sovereign cloud environments in critical infrastructure deployments face increased adversarial threats across software delivery pipelines and air gapped operations. The challenges are further compounded by terrestrial and non terrestrial network (TN-NTN) segments. This systematic literature review (SLR) investigates the intersection of secure software delivery, Artificial Intelligence (AI) vulnerability triage, and DevSecOps within the constraints of TN-NTN and air gaps. For this survey, we included 22 studies identified through a two-arm PRISMA 2020 systematic literature review, which originated from 429 database records and 6 snowball citations (2 anchor citations). Additionally, we propose a taxonomy that maps the included studies against a two dimensional framework: (i) AI complexity: classical Machine Learning (ML), hybrid, Generative Artificial Intelligence (GenAI)/ Large Language Models (LLMs) and (ii) connectivity assumptions: air gap, partial non terrestrial (NTN), terrestrial. The analysis of results and outcomes is presented to illustrate research trends in AI-driven vulnerability triage in connectivity-constrained environments. Finally, we offer future research directions and identify open challenges for readers interested in exploring this area further.
3D human pose estimation is a fundamental research task in computer vision, with the goal of predicting the 3D coordinates of key human joints to construct a kinematic skeleton. Given the robustness of existing 2D pose detectors, lifting multi-frame 2D keypoint sequences to 3D poses has become the dominant paradigm in the field, as this approach significantly improves estimation accuracy compared to processing single frames. Although the Transformer architecture excels at processing sequential data, existing models often lack physical priors of human structure, requiring vast amounts of data for learning and being prone to generating anatomically implausible poses. Furthermore, the global self-attention mechanism can introduce computational redundancy and noise interference. To address these issues, this work proposes a new model named ADSGFormer, which employs a novel dual-stream architecture to process spatial and temporal information in parallel. It leverages complementary information through a fusion mechanism to eliminate ambiguities, thereby enhancing the model’s representational capability and estimation accuracy. At the core of the model is an innovative temporal attention module that more effectively captures both global and local dependencies and effectively utilizes prior knowledge. Experimental results on Human3.6M and MPI-INF-3DHP show that ADSGFormer achieves competitive state-of-the-art performance without relying on multi-hypothesis sampling or diffusion-based post-processing.
Tor is a popular low-latency anonymous communication system that protects user privacy via layered encryption. However, Website Fingerprinting (WF) attacks can still infer browsing behavior by analyzing statistical patterns in encrypted traffic. Existing multi-tab WF methods primarily rely on temporal features and employ fixed attention windows that treat all traffic segments uniformly, ignoring the fact that multi-tab traffic exhibits varying degrees of overlap across different time periods. To address these limitations, we propose CRAFT, a complexity-routed attention framework with frequency-temporal fusion for multi-tab website fingerprinting. First, we introduce a dual-branch feature extractor that combines a temporal convolutional backbone with a frequency-domain branch for capturing periodic traffic rhythms, fused via a learned gating mechanism. Second, we design a complexity-aware adaptive routed attention mechanism, where a lightweight router predicts per-token traffic mixing complexity and dynamically adjusts the attention scope through a Gaussian distance bias, focusing locally on simple single-site segments while expanding to broader context in complex overlapping regions. Extensive experiments across multiple multi-tab settings demonstrate that CRAFT achieves strong and consistent improvements over existing approaches.
For deep-learning-based AI systems, copyright risks are now pervasive across the entire lifecycle, rather than being confined to isolated artifacts such as data, models, or outputs. While intrusive audit approaches that modify training pipelines such as watermarking have been actively studied, they require intervention during model training and may harm model performance. This survey focuses on non-intrusive auditing methods based on mechanisms that do not intervene in the state of the involved copyright assets themselves. Specifically, we organize existing work according to the lifecycle stages that these non-intrusive audits reason about, from training data to trained models and downstream generated content. This lifecycle-oriented view reveals four main copyright auditing tasks: (1) data usage audit, (2) model derivation audit, (3) model usage audit, and (4) data derivation audit. For each task, we provide a structured review of representative approaches and discuss their core contributions. Across these tasks, existing methods can be further categorized into two complementary methodological families: statistical auditing, which estimates the likelihood of copyright-relevant misuse through measurable statistical discrepancies, and evidentiary auditing, which constructs structured technical evidence that may support further assessment of copyright claims. Overall, this survey provides a structured framework for understanding and interpreting non-intrusive copyright auditing methods for different auditing tasks across the AI system lifecycle.
While Large Language Models (LLMs) offer transformative potential across domains, often outperforming human benchmarks in various tasks, they remain vulnerable to exploitation by users aiming to override their safety protocols. Despite the progress achieved through red teaming methodologies in uncovering and mitigating such vulnerabilities, one notably persistent technique, referred to here as the Goofy Game, which leverages role-playing strategies, continues to bypass many existing safeguards. This technique can elicit unsafe responses from LLMs, which, although seemingly benign in isolation, could lead to severe consequences when deployed within high-stakes environments such as clinical decision-making or patient communication. In this study, we build on the insights from our previous exploratory experiments and analyse how a malicious user, even without technical knowledge of the internal architecture and parameters of generative AI models, could create a role-playing prompt that coerces a language model (LLM) into generating incorrect and potentially harmful clinical suggestions. Our objective is to elucidate a particular vulnerability scenario and provide insights that will contribute to future advancements in the development of secure and reliable AI systems.
The precise description of salient objects and spatial relationships in remote sensing images is crucial for the remote sensing image captioning (RSIC) task. Nevertheless, due to limited integration of local and global features, existing methods often face challenges in distinguishing visually similar objects and capturing meaningful scene context. To address this issue, it is necessary to enhance the feature representations of remote sensing images, enabling more precise and semantically coherent caption generation. Therefore, this paper presents a comprehensive investigation into global and local feature extraction and fusion for remote sensing images. We propose a remote sensing image captioning model based on Sparse Mechanism-guided Feature Enhancement (SMFE). Specifically, the model employs Oriented R-CNN and CLIP to extract local and global features from remote sensing images, respectively. These features are then fused through a proposal-slot gated aggregation module, which uses depthwise separable convolutions with different dilation rates for lightweight token-slot mixing on a fixed proposal-slot layout. To further refine the visual representation, a learnable sparse gate is incorporated into the encoder to suppress redundant attention responses. Its main effect is feature selection and noise suppression, rather than direct acceleration. Finally, a Transformer decoder is used to generate descriptions. Experiments on RSICD, UCM-Captions, Sydney-Captions, and VRSBench-Captions show that SMFE is effective on conventional RSIC benchmarks. The complexity analysis also shows that the full pipeline introduces extra cost because it includes raw-image visual encoders. The VRSBench results show that large VLMs remain stronger on most metrics, which clarifies the current scope of the method.
Smart meters, utilizing their large-scale distributed deployment architecture, facilitate real-time acquisition of fine-grained residential load data, which significantly enhances the accuracy of Short-term Residential electrical Load Forecasting (SRLF) models. However, SRLF typically requires fine-grained user data as input, which raises privacy concerns. Moreover, smart meters, as computing resource-constrained devices, are unable to perform complex training processes locally. Furthermore, the large-scale deployment of smart meters imposes significant communication overhead and latency on networks. To address these challenges, we propose an end-edge-cloud hierarchical Federated Learning (FL) framework that leverages fine-grained residential load data from numerous smart meters in a privacy-preserving and resource-efficient manner. Specifically, the primary training and prediction tasks are shifted to edge servers proximate to the smart meters. An inner-product functional encryption scheme is employed, enabling local training and prediction to be performed by the edge server for multiple smart meters without their raw data being accessed. Furthermore, a hierarchical aggregation strategy ensures the equitable and comprehensive utilization of distributed data. Extensive experiments demonstrate our scheme achieving better performance in the practical environment.
Accurate and real-time detection of projectile impact points over sea surfaces is critical for modern maritime weapon testing and fire-control adjustment. How-ever, splash signatures generated at impact are extremely short-lived, often visible for only 1–3 consecutive radar image frames, and are easily submerged in high sea-state-induced, rapidly fluctuating radar backscatter caused by wind-driven waves and surface roughness. These characteristics severely limit the reliability of conventional detection and modeling approaches, particularly under complex sea-state conditions. To address these challenges, this study proposes a domain-adaptive autoencoder-based reconstruction framework for real-time splash detec-tion in maritime radar video. An amplitude-alignment preprocessing module is introduced to alleviate cross-sea-state intensity inconsistency and stabilize feature representation across varying sea conditions. The framework further incorporates a dual-evidence detection mechanism that jointly leverages clutter–background re- construction residuals and short-term splash dynamics learned from limited splash samples. In addition, a cascaded three-stage detection architecture is designed to organize detection from coarse to fine, facilitating efficient and structured splash localization in radar video streams. Experiments on real maritime radar sequences indicate that the proposed framework enables reliable real-time splash detection under diverse sea states, including severe clutter, frame loss, and dynamic inter-ference.
Artificial Intelligence (AI) systems used in sensitive domains must balance privacy protection and explainability. Differential Privacy (DP) provides formal privacy guarantees, while Explainable AI (XAI) aims to make model decisions understandable. However, privacy-preserving noise often degrades explanation quality, making their combination challenging. This survey provides a structured review of literature published between 2020 and 2025 and differs from prior threat-oriented or benchmark-driven surveys by adopting a mechanism-centered perspective that explicitly examines how DP techniques interact with different explanation paradigms. From this broader technical and system design perspective, we analyze DP and XAI interactions beyond purely threat-oriented considerations. We analyze these interactions through the lens of the privacy-explainability-utility trade-off, which shows that improving privacy can harm either model performance or explanation quality. To help organize and advance this growing field, we propose a five-category taxonomy that characterizes how DP and XAI connect in machine learning systems. We also introduce the Differential Privacy-Explainable AI (DP-XAI) interaction matrix, which maps combinations of privacy techniques and explanation methods that have been studied. Our analysis identifies several research gaps, especially in concept-based explanations and underused architectures, and highlights concrete directions for future research on building machine learning systems that are both privacy-protected and meaningfully explainable.
As wireless communication systems advance toward the sixth generation (6G), intensive research and development efforts are underway. To identify enabling technologies that meet the unprecedented requirements of next-generation networks, including ubiquitous connectivity, sustainability, ultra-high throughput, and near-zero latency. To fulfill these ambitious goals, emerging paradigms including Digital Twin (DT), High Altitude Platform Systems (HAPS), terahertz and petahertz communications, Internet of Things (IoTs), Integrated Sensing and Communication (ISAC), Orthogonal Time Frequency Space (OTFS) modulation, Network Slicing (NS), and Quantum Communication (QC) are being integrated with Cell-Free (CF) network architecture. Although CF networks offer significant advantages in coverage and capacity, their distributed nature introduces several challenges, including antenna coordination, resource allocation, channel estimation, security, task offloading, and computational complexity. Recent advances in Artificial Intelligence (AI), particularly Deep Learning (DL), present promising solutions to mitigate these issues to cater global coverage with seamless connectivity and enhanced security. This paper provides a comprehensive review of state-of-the-art technologies synergized with CF communications and explores DL-driven methodologies to realise the performance targets of future 6G wireless networks.
Workflow agents that read messages, retrieve documents, and take follow-up actions can improve productivity, but they also widen the security boundary by combining untrusted content with permissions to access and modify external resources. This issue is especially important in safety-critical domains such as healthcare, where inappropriate access to patient-related information, wrong-recipient disclosure of reports, or deletion of clinically important workflow messages can have serious consequences. We present a controlled framework for evaluating risk and auditability in workflow agents under varying permission scopes and policy settings. Our main testbed is a fully offline, deterministic, and reproducible simulator that isolates how broader permissions amplify harm under permissive settings and how strict policy gating shifts blocking from scope restriction to semantic enforcement. To provide a limited external-validity check, we also add a lightweight real-agent validation layer using a local tool-using language model over sandbox workflow cases built from curated real radiology reports from MIMIC-IV-Note. Across both layers, permissive broad access enables harmful forwarding or deletion behaviors, while stricter controls reduce realized harm through policy-based or permission-based blocking. We further implement a practical tamper-evident audit mechanism based on a hash-chained append-only log and show that post-hoc modifications of execution history are detectable and localizable to the first failing block. Together, these results support a layered view of workflow-agent safety in which least privilege, semantic policy gating, and verifiable execution records play complementary roles.
Post-disaster search and rescue missions have extremely high requirements for timeliness and resource utilization efficiency. Owing to their flexibility, increasingly diverse remote sensing payloads, and rapid deployment capabilities, Unmanned Aerial Vehicles (UAVs) have become an important part of post-disaster rescue efforts. However, the capabilities of a single UAV are limited, and a multi-UAV collaborative system is required to improve mission effectiveness, among which the task allocation strategy is the key factor affecting the overall performance. Previous studies have not fully considered the dynamic coupling relationship between UAV payload and battery capacity, nor have they collaboratively optimized material and battery resources during the allocation process, which limits the improvement of the mission completion rate. In this study, we propose a Dynamic Task Allocation algorithm for Search and Rescue (DTASR). First, a system model is established to quantify the dynamic impact of the load on power consumption and flight range. Second, a two-stage allocation mechanism for search and collaborative rescue is designed, and a multi-objective optimization problem is balanced by introducing a multi-value function. A collaborative material allocation strategy based on the remaining battery capacity is proposed to achieve the efficient collaborative utilization of battery power and materials. The simulation results show that DTASR can maintain a stable mission completion rate under different scenario scales, and its performance is significantly better than those of the existing mainstream algorithms. These results verify the effectiveness and robustness of the algorithm in dynamic post-disaster environments.
This paper investigates the max-min fairness problem in six-Dimensional Movable Antenna (6DMA) enhanced Non-Orthogonal Multiple Access (NOMA) scenarios, where multiple users adopt adaptive Successive Interference Cancellation (SIC) to mitigate inter-user interference. To maximize the minimum achievable rate among users, the joint optimization of 6DMA configuration and beamforming of the BS as well as the adaptive SIC decoding of users is formulated as a high-dimensional mixed-integer nonconvex problem. We decompose this challenging problem into subproblems of continuous and discrete variables, then propose a two-layer optimization framework. In the inner layer, the transmit beamforming and 6DMA configuration are alternately optimized. In the outer layer, a gene archive Assisted Elitist Genetic Algorithm (AE-GA) is proposed to optimize the adaptive SIC decoding. Theoretical analysis demonstrates that the proposed optimization framework always converges to a stable solution with polynomial complexity. Simulation results verify the fairness enhancement of our proposed scheme over SDMA and conventional NOMA.
Constrained by edge computing resources and model capacity, embodied agents often struggle to maintain high inference accuracy in complex dynamic scenarios when performing real-time visual perception tasks. Although offloading difficult samples to the cloud for collaborative inference is an effective way to mitigate this problem, existing solutions generally overlook the challenges posed by network transmission uncertainty to inference timeliness. The traditional network layer lacks the ability to perceive the semantics of application-layer visual tasks and cannot distinguish the urgency of inference requests. Consequently, high-risk difficult samples compete for bandwidth with ordinary traffic in congested environments, undermining the reliability of embodied intelligence systems. To address these challenges, this paper proposes a Reliable Application-aware Collaborative Inference (RACI) framework, which establishes a cross-layer mechanism linking application semantics to network scheduling. The core idea of RACI is to bridge the gap between the application layer and the network layer: At the application layer, an adaptive decision mechanism based on Scene Ambiguity Score (SAS) is designed to identify difficult samples requiring offloading with low computational overhead by jointly measuring edge-side inference uncertainty and scene object density; at the network layer, the offloading decision semantics are encoded into the Flow Label field of IPv6 packets, enabling network devices to identify the urgency of inference traffic and perform differentiated flow scheduling. Experiments in object detection scenarios show that RACI achieves detection accuracy improvements in scenarios where edge model capabilities are insufficient, and effectively supports low-latency transmission of critical inference traffic under network congestion.
This paper systematically investigates the security, privacy, and ethical risks, as well as the traceability challenges of OpenClaw, a locally executable AI agent system for natural language interaction and real-world task completion. While OpenClaw shows strong potential for personal assistance, office automation, cross-platform task management, and information integration, it also raises serious security, privacy, and ethical concerns. By analyzing its system architecture, core functionalities, deployment model, and representative application scenarios, this paper aims to reveal the risks that may arise when such a highly privileged agent is integrated into personal and organizational digital environments. We focus in particular on the challenges associated with persistent local storage, tool invocation, cross-context information aggregation, multi-user interaction, and the integration of plugins and external services. We argue that these issues constitute major barriers to the trustworthy deployment and widespread adoption of this technology. Finally, we summarize the open challenges in security defenses, privacy protection, ethical governance, and traceability in agent use, and call for joint efforts from researchers, developers, deployers, and regulators to build AI agent systems that are safer, more reliable, and more trustworthy.
Generative AI systems are increasingly used not only to produce content but also to retrieve data, invoke tools, and execute actions. This work examines the security and safety implications of that shift across content-level, model-level, and agentic threats. We analyze how attacker access requirements, system autonomy, and the scope of potential harm change as models move from generating artifacts to executing operations through tool chains and external APIs. We then assess technical countermeasures including detection, watermarking, alignment, and emerging agentic safeguards, and show that several depend on forms of institutional coordination that current governance arrangements do not yet provide. Across the cases examined, capability deployment and attack-surface expansion repeatedly outpace defensive responses as systems move from generating content to executing real-world actions.
While micro-Doppler signatures are a proven modality for discriminating between drones and birds, their reliability remains questionable in low-SNR, data-constrained environments where other methods such as deep learning models might struggle. This paper presents a systematic analysis of ten statistical and physics-motivated handcrafted features for micro-Doppler classification under controlled signal degradation, using a publicly available 77 GHz frequency modulated continuous wave (FMCW) radar dataset. Micro-Doppler spectrograms are corrupted using Additive White Gaussian Noise (AWGN), phase noise, and their combined effects across a range of −10 dB to 10 dB signal-to-noise ratios (SNRs) and 1°- 10° phase noise levels. Features extracted from the raw and corrupted spectrograms are evaluated via performance metrics within a 5-outer × 5-inner stratified cross-validation framework for a Support Vector Machine (SVM) and Random Forest classifier, with fixed hyperparameters across all noise levels chosen through grid search. Feature relevance under noise is understood through the lens of permutation based importance calculated as the decrease in macro-averaged F1 score induced by random shuffling of a single feature while keeping all other features fixed. For noise-free data mean classification accuracies of 0.916 ± 0.095 for the SVM and 0.916 ± 0.060 for the Random Forest classifier were achieved with F1 scores of 0.909 and 0.912. Experimental results reveal that entropy-based and side-lobe features maintain stable discriminative performance under severe noise, reporting macro-averaged F1 scores of 0.773 and 0.831 for the SVM and Random Forest, respectively. Feature-wise evaluation and permutation-based importance analysis reveals that certain features retain complementary discriminative power even when their standalone importance appears low. These findings highlight the importance of principled feature design and provide insight into feature vulnerability and resilience, offering guidance towards relatively noise-resilient and interpretable feature selection for next-generation radar classification systems.
