The increasing interconnection of smart-grid operational technology, industrial-edge services, and utility information systems creates a critical need for resilient and continuously monitored industrial cyber-physical communication. Although quantum key distribution (QKD) can strengthen session-key establishment for advanced metering infrastructure, distributed energy resources, substation automation, supervisory control, and utility-core services, practical QKD deployments remain vulnerable to implementation-level side-channel attacks that can compromise the cryptographic protection layer without directly targeting conventional network packets. This paper presents a QKD-secured industrial smart-grid cyber-physical system framework for simulating and detecting adaptive side-channel attacks. The proposed 36-node industrial communication architecture integrates AMI devices, DER controllers, PMU and substation automation components, industrial-edge gateways, QKD modules, key-management services, SCADA and utility-core servers, security-operation-center components, and adversarial access points. A 100,000-record cyber-quantum dataset is generated across 12 operating conditions comprising normal communication and 11 adaptive QKD side-channel attacks: detector blinding, time shift, wavelength switching, Trojan-horse probing, photon-number splitting, decoy-state spoofing, RNG bias, calibration manipulation, local-oscillator manipulation, synchronization spoofing, and combined adaptive quantum hacking. Each scenario introduces coupled primary and secondary perturbations across optical, detector, timing, synchronization, randomness, calibration, photon-statistical, leakage, key-generation, encryption, and industrial-network-performance features. To support intelligent industrial security monitoring, the proposed Quantum-aware Mamba–Kolmogorov–Arnold Network (Q-MambaKAN) organizes device, network, QKD, side-channel, encryption, and risk evidence into an ordered cyber-quantum representation processed through selective state-space learning, side-channel attention, nonlinear KAN mapping, adaptive fusion, and multi-task prediction heads. Results show that the QBER increases from 0.071 during normal operation to 0.426 under combined adaptive quantum hacking, while encryption success decreases from 98.1% to 0%. Q-MambaKAN achieves a 99.48% binary detection accuracy, a 99.70% binary F1-score, a 97.60% multiclass macro-F1, and a risk RMSE of 0.021.
This paper presents a comparative benchmarking study evaluating the regulatory comprehension of four open-source large language models, TinyLlama-1.1B-Chat, Gemma-2B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, on the EU Artificial Intelligence Act (Regulation EU 2024/1689). A custom benchmark of 100 multiple-choice questions was constructed and classified across three dimensions, namely cognitive difficulty, knowledge type, and regulatory mechanism, with a balanced answer key distribution. Each model was evaluated under vanilla inference, relying exclusively on parametric knowledge, and retrieval-augmented generation (RAG), in which relevant Act passages are retrieved using a FAISS-indexed sentence embedding pipeline with standardized top three chunk retrieval applied uniformly across all models. Vanilla accuracy ranges from 22.0% for TinyLlama-1.1B to 80.0% for Llama-3.1-8B, indicating that larger models perform better. Under the retrieval configuration employed, RAG improves performance of the evaluated models, with gains from 3.0 percentage points for TinyLlama-1.1B to 17.0 for Gemma-2B and 10.0 for both 7B–8B models. Questions related to procedural knowledge, governance and enforcement emerge as weaknesses in the baseline, vanilla setting, partially mitigated by retrieval. Qualitative analysis identifies two RAG failure modes: retrieval failures, where the embedding mechanism returns informationally insufficient passages, and integration failures, where the correct passage is retrieved but not correctly exploited. The study contributes an empirical characterization of small and medium language model regulatory comprehension and a reusable 100-question benchmark.
Zero-trust security continuously re-evaluates the trustworthiness of industrial devices and reacts by rerouting, isolating, or rescheduling traffic. In a time-sensitive network (TSN) that carries feedback control loops, however, every such reaction is itself a control-plane disturbance: a reroute that meets every deadline can still deliver stale measurements, and an optimizer that crashes mid-reconfiguration can leave the network in an undefined state. This paper presents ZTSafe, a scheduling architecture that treats physical safety—not attack blocking—as the object of guarantee. The guarantee has two distinct layers: compliance with the communication contract yields a deterministic invariance result conditional on the stated plant, disturbance, synchronization, and trusted-base assumptions, whereas the risk bound’s 1−δ coverage is an empirical probabilistic calibration result. ZTSafe (i) synthesizes, offline and per control loop, a communication safety contract that bounds delay, age of information (AoI), consecutive losses, jitter, and path risk such that the physical state remains in its safe set under those assumptions; (ii) converts zero-trust evidence into conservative risk upper bounds and couples the admissible path-risk budget to the runtime safety margin of the plant; and (iii) places the scheduling optimizer outside the trusted computing base: an independent runtime shield checks every proposed schedule against the contracts, and on solver timeout, crash, or infeasibility the system atomically switches to a pre-checked fallback instead of executing an unverified approximate solution. Here, “verified” means independently checked by the shield, not machine-verified; a systematic shield defect or compromise of the remaining trusted computing base voids the deterministic claim. On a hardware TSN testbed with three physical control loops and fourteen attack and fault scenarios, ZTSafe reduces safe-set violations by 92.9% relative to the strongest baseline (12.8% to 0.9%; two-proportion z=39.4, p<10−15) while sustaining 94.3% on-time completion of critical traffic, recovers within three control periods, and executes zero unverified configurations across 10,000 injected solver failures.
LoRaWAN has become a widely adopted communication technology for Internet of Things applications because of its long-range communication capability and low power consumption. However, resource-constrained LoRaWAN networks remain vulnerable to physical-layer attacks such as jamming and battery-depletion attacks. This paper investigates architectural compression and Knowledge Distillation (KD) for lightweight multi-attribute Physical Layer Authentication (PLA) in LoRaWAN networks. A higher-capacity Teacher model is used to train progressively smaller Student models, and three Student-training strategies—Direct training, Hard-KD, and Soft-KD—are systematically compared to determine whether Teacher supervision provides an advantage beyond architectural compression alone. The framework uses the Received Signal Strength Indicator (RSSI), altitude, and the battery level as input attributes and is evaluated offline using a device-aware semi-synthetic dataset derived from the publicly available Brno LoRaWAN dataset. The resulting dataset contains Normal Communication, Jamming Attack, and Battery Depletion Attack traffic. The experimental results show that Student (8) preserves a classification performance close to that of the Teacher across all three training strategies, achieving Macro F1-scores of 90.619 ± 0.056% for Direct training, 90.591 ± 0.069% for Hard-KD, and 90.626 ± 0.083% for Soft-KD, compared with 91.007% for the Teacher. Student (8) reduces the number of trainable parameters from 2435 to 59 (97.6%) and the serialised model size from 46.65 KB to 2.80 KB (94.0%), while exhibiting a substantially higher measured batch-processing throughput under the evaluated CPU-based implementation. A feature-ablation analysis further indicates that the RSSI and battery level provide the dominant discriminatory information, while the altitude provides a smaller but measurable contextual contribution. The comparable performance of Direct, Hard-KD, and Soft-KD demonstrates that KD does not consistently outperform direct training for the evaluated Student architectures; rather, architectural compression itself accounts for much of the retained performance. These findings demonstrate that compact Student architectures can substantially reduce the model complexity while preserving a strong classification performance within the evaluated multi-attribute PLA framework, supporting the further investigation of lightweight network-side security mechanisms for LoRaWAN deployments.
Large language models (LLMs) are increasingly being deployed as multi-agent systems (MASs), yet limited evidence exists on when such architectures are warranted and how reliably they perform in decision-critical workflows. This paper investigates these questions through veterinary insurance claims adjudication, a document-intensive, policy-driven task requiring structured reasoning, regulatory compliance, and human oversight. We first propose a four-condition framework for assessing MAS suitability and identify claims adjudication as the highest-priority veterinary application. We then develop a self-hosted MAS integrating specialised LLM agents with deterministic policy evaluation, immutable audit logging, and mandatory human review. To support reproducible evaluation, a synthetic benchmark is released, comprising 150 veterinary insurance claims with independently computed ground-truth adjudication outcomes. The proposed system achieves a 97.3% workflow completion rate and 90.0% adjudication accuracy, increasing to 92.5% for completed claims and approaching the corpus ceiling of 96.0%. The evaluation shows that reliable deployment depends primarily on system engineering rather than model capability alone. Explicit orchestration and verification eliminate common LLM tool-use failures, deterministic adjudication achieves perfect agreement with the reference oracle, and completing the inter-agent policy contract increases agreement from 91.8% to 100%. These results demonstrate that hybrid MASs combining probabilistic document understanding with deterministic policy execution provide a practical and auditable approach for decision-support applications in insurance and other policy-governed domains.
As deep neural networks continue to scale and enable emerging applications such as agentic AI systems, training increasingly relies on distributed paradigms across heterogeneous edge devices. However, this shift introduces significant security challenges, particularly model poisoning attacks, which are largely underexplored in model-parallel settings. To address these challenges, we propose a trusted and attack-resilient mechanism for distributed DNN training that supports both data and model parallelism. The mechanism leverages a blockchain-enabled infrastructure to ensure the tamper-resistant and auditable execution of security-critical operations. It introduces a Loss-aware Credit Evaluation mechanism to assess agent reliability based on group-level training dynamics and a Shuffling-based Isolation Mechanism to progressively cluster and isolate malicious agents across training epochs. In addition, Byzantine-tolerant aggregation (BTA) is employed to further mitigate adversarial influence during model aggregation. Extensive experiments demonstrate that the proposed mechanism achieves superior robustness and efficiency compared with state-of-the-art methods under diverse poisoning attack scenarios.
Immersive technologies—including virtual reality (VR), augmented reality (AR), mixed reality (MR), extended reality (XR), and the metaverse—have undergone significant development over the past decade and are increasingly being integrated into a broad range of application domains [...]
Currently, in high-density indoor environments such as businesses and factories, managing human presence and access control remains a significant challenge. Traditional access control systems based on facial recognition or card scanning typically only record authentication events at the point of entry. Therefore, continuous monitoring, presence detection, or restricted area surveillance are limited. This paper proposes an integrated indoor person detection and management framework based on centralized multi-camera processing for deployment-oriented identification surveillance. The proposed framework combines SCRFD and ArcFace to perform enrollment-based face recognition and distinguish between enrolled and unknown identities. During the experimental evaluation, the facial recognition module using the SCRFD 2.5G configuration achieved a recognition accuracy of approximately 88.2%. YOLOv11n is integrated with DeepSORT to detect and continuously track individuals within the monitored area. Experimental results showed the system achieving an average processing performance of 10.6 FPS, demonstrating the feasibility of the proposed architecture for small- to medium-scale indoor surveillance applications. Additionally, this system framework integrates event-driven spatial analysis using virtual boundaries and surveillance zones to support entry/exit counting, presence monitoring, and intrusion detection in restricted areas. Experimental results demonstrate that the proposed system framework provides consistent identification monitoring performance, stable multi-object tracking capabilities, and an efficient event management mechanism for typical indoor surveillance scenarios. This work offers a centralized, deployment-oriented surveillance architecture suitable for practical indoor security management and access control applications utilizing multiple cameras.
With the rapid development of Industrial Internet-oriented smart energy systems, hydro-turbine generator units are increasingly monitored through networked sensors, industrial communication infrastructures, and edge/cloud-based condition-monitoring platforms. These Internet-connected monitoring environments provide abundant vibration data for intelligent operation and maintenance (O&M) but also introduce a challenging unsupervised cross-scenario diagnosis problem. Specifically, diagnostic models trained on labeled historical data may suffer severe performance degradation when deployed to unlabeled online data collected under different hydraulic conditions, rotational speeds, or operating conditions. Furthermore, existing domain adaptation methods, in their pursuit of distribution alignment, frequently overlook a critical bottleneck that limits generalization performance: inter-class entanglement. Specifically, under intense hydraulic background noise and cross-condition distribution shifts, features belonging to distinct fault types are highly susceptible to aliasing within the feature space. To overcome these issues, this paper proposes a Conditional Wasserstein Adversarial Network with Bi-level Prototype Disentanglement Regularization (CWAN-BPDR). First, a Conditional Wasserstein Adversarial Network (CWAN) is constructed by combining the smooth-gradient property of Wasserstein distance with conditional adversarial alignment, thereby achieving stable and fine-grained category-level domain adaptation. Furthermore, to alleviate the inter-class entanglement problem that may arise during cross-domain alignment, a Bi-level Prototype Disentanglement Regularization (BPDR) term is designed. By jointly implementing source–target prototype alignment and prototype–feature bidirectional alignment, BPDR explicitly suppresses inter-class confusion and enhances intra-class compactness and inter-class separability in the feature space. Experimental results on the JNU and NEFU datasets demonstrate that CWAN-BPDR achieves average diagnostic accuracies of 97.82% and 98.99%, respectively, while significantly mitigating label entanglement in challenging cross-operating-condition tasks. These results indicate that the proposed method can effectively transfer diagnostic knowledge acquired from labeled historical operating conditions to unlabeled online monitoring data. It can therefore serve as an offline-trained diagnostic module for Industrial Internet of Things-based condition-monitoring platforms in hydropower systems.
Arabic question-answering systems in high-sensitivity domains require not only accurate retrieval but also reliable evidence grounding and effective hallucination mitigation, as incorrect or unsupported responses may have serious consequences. Existing retrieval and generation approaches do not fully integrate reliable lexical retrieval, semantic understanding, and evidence-based answer verification within a unified framework for these domains. To address this limitation, this study proposes an Enhanced Hybrid Retrieval-Augmented model that combines BM25-based lexical retrieval, dense semantic scoring, semi-structured metadata, domain-aware classification, source-trust and freshness indicators, and evidence-based answer verification. The model was evaluated on a corpus of 128,297 Arabic documents using 200 expert-validated questions and compared against BM25, Dense, and Classical Hybrid retrieval configurations. Expert assessment showed that the Enhanced Hybrid Model produced 198 fully grounded correct answers out of 200, achieving a grounded-correctness rate of 99.0%, compared with 96.5% for Classical Hybrid, 87.0% for BM25, and 86.5% for Dense. It also achieved the strongest answer-level Hit@5 performance, the best golden-answer ordering, and the highest dynamic mixture-based F1-score. These findings demonstrate that integrating reliable retrieval, metadata-aware ranking, and evidence-grounded answer verification can substantially improve the reliability and trustworthiness of Arabic question-answering systems in high-sensitivity domains, providing a practical foundation for future evidence-based intelligent information systems.
Unwanted telemarketing calls are a persistent source of consumer frustration and a legally regulated issue in Germany. At first glance, the idea of addressing this problem with AI-based voice technology appears contradictory: why should an automated caller help restore trust in a communication channel that has been damaged by aggressive outbound practices? This design-oriented case and prototype study argues that the paradox can be resolved through a different design logic. Rather than using AI to intensify persuasion, we present a role-based voice-agent architecture that constrains conversational behavior through narrow task boundaries, explicit escalation rules, and auditable data handling. The paper reports a transfer project involving FH Aachen students, Capgemini, and Fairdient GmbH. Methodologically, the work is positioned as a design-oriented case study with a prototype artifact. The contribution is threefold: first, we describe a three-agent architecture for outbound screening, consent-aware explanation, and inbound service; second, we derive governance principles for legally and ethically sensitive telephony, including transparency, bounded knowledge, privacy-preserving deployment, and human fallback; and third, we propose an evaluation framework covering conversion, compliance, hallucination control, user trust, and cost per validated outcome. The prototype does not yet claim large-scale field effectiveness. Instead, it offers a structured and empirically testable design for trustworthy voice automation in a domain where misuse, opacity, and user distrust are especially pronounced.
Industrial IoT maintenance often stops at health prediction, leaving maintenance, rescheduling, safety, and communication to separate decision processes. This study presents MIRA, a safety-constrained graph-based multi-agent reinforcement learning architecture for joint prescriptive maintenance, production rescheduling, and event-triggered communication. Machine condition was estimated from CNC milling data using temporal convolutional models; because predictive uncertainty failed a predefined validation gate, the controller used deterministic health estimates. Evaluation covered five controllers, six simulated scenarios, and 1800 matched episodes. Relative to Graph-MAPPO, MIRA reduced operational cost by 9.38%, weighted tardiness by 28.10%, unexpected failures by 17.39%, message count by 84.98%, and transmitted data by 83.83%, while increasing on-time completion by 23.55%, without a detectable difference in corrected critical-message recall. Across the three independently trained seeds, failures, safety violations, and message count favored MIRA consistently, whereas cost and tardiness favored MIRA in two seeds. Disabling the execution shield increased safety violations from 0 to 3.56 per episode. Post-training variation in the projected-health safe-start threshold from 0.124 to 0.132 produced no safety violations and only small changes in aggregate operational outcomes. Cross-domain health transfer to PHM 2010 failed without adaptation. The results support simulator-level decision coordination, while broader replication, variable-size deployment, and factory validation remain necessary.
Machine Learning (ML)-based malware detectors perform exceptionally well on standard benchmarks, but their robustness to evasive malware remains fragile. On EMBER2024, strong tree-based models achieve high accuracy on clean test sets yet miss a large number of evasive Windows PE samples. Through SHAP-based explainability, matched error comparisons, feature-family analysis, and margin diagnostics on the validation split, we identify consistent attribution and decision-margin differences between baseline true negatives and false negatives. We further show that false negatives occur at different decision depths, including both near-boundary cases and samples misclassified with high benign confidence. We propose C-GUARD, a context-adaptive conformal gated detector that preserves the optimized high-performing baseline while selectively invoking an auxiliary rescue detector on a targeted subset of baseline negative predictions. C-GUARD combines out-of-fold rescue learning to distinguish baseline false negatives from true negatives with adaptive gating under an explicit false-positive budget. A one-sided intervention rule preserves all baseline malware decisions. On the EMBER2024 standard test set, C-GUARD recovers 178 additional malware samples at the cost of 82 additional false positives. On the evasive challenge set, it recovers 17 additional evasive malware samples.
Telemedicine in emergency and remote care relays medical images from ambulances and rural clinics to a hospital edge-computing server over a congested wireless uplink. Existing work prices such transmissions per bit or per quality-of-experience; neither metric captures the clinical value of a medical transmission. Diagnostic utility vanishes below a modality-specific acceptability floor rather than degrading gracefully, the deadline is determined by triage acuity rather than by the network, and a missed finding is far costlier than a false alarm. A per-bit clearing price therefore disadvantages the node that has expended local compute to produce a compact, diagnostically sufficient stream. We propose SemAuc, a triage-aware truthful mechanism for medical-image admission over a rate-splitting uplink, in which the shared semantic knowledge base rides the common stream, and case-specific residuals ride private streams. SemAuc filters tiers below the diagnostic floor and beyond the clinical deadline, reserves a regulated-price lane for life-threatening cases, and allocates remaining capacity through a single-parameter contestable auction whose bid-independent pre-selection step satisfies the conditions of Myerson’s lemma. The contestable lane is dominant-strategy truthful, individually rational, near-linear in the number of nodes, and achieves a constant-factor density-greedy welfare guarantee; the clinical lanes follow from triage policy without disturbing these properties. Diagnostic value is grounded by an offline kernel fitted on BraTS and CheXpert. On a Rayleigh-faded uplink at two hundred contending nodes, SemAuc preserves the high-acuity diagnostic service-level objective where bit-centric benchmarks fail, and tracks the offline optimum.
The rapid growth of cloud and distributed computing systems has increased the complexity of real-time request distribution under dynamic and multi-factor conditions. In such environments, load-balancing decisions must simultaneously consider uncertain and interdependent parameters, including server load, response time, and resource capacity. Fuzzy logic is an effective tool for modeling such uncertainty; however, the expansion of linguistic variables often leads to a rule-explosion problem, which increases computational complexity and reduces the real-time applicability of fuzzy load-balancing systems. This study proposes an explainable multi-objective quantum-inspired fuzzy optimization approach for scalable load balancing in complex computing environments. The proposed model integrates fuzzy inference with a Grover-inspired classical search strategy to optimize the selection of fuzzy rule subsets. The Grover-inspired component is implemented as a classical simulation rather than a gate-based quantum circuit. A multi-objective evaluation function is formulated to jointly assess rule accuracy, coverage, interpretability, and compactness. This formulation enables the model to reduce redundant fuzzy rules while preserving decision transparency and maintaining reliable load distribution performance. The proposed approach is evaluated in a simulated cloud computing environment with heterogeneous servers and dynamic request arrival patterns. Comparative experiments are conducted against classical load-balancing strategies, conventional fuzzy load balancing, and evolutionary fuzzy optimization methods, including GA-FLB and PSO-FLB. The experimental results show that the proposed model reduces the size of the fuzzy rule base while maintaining competitive response time, load distribution quality, SLA compliance, and decision interpretability. These findings indicate that the integration of Grover-inspired classical search mechanisms with fuzzy reasoning provides a promising direction for developing scalable, compact, and explainable load-balancing models for next-generation intelligent computing systems.
The disruptive potential of AI agents is becoming increasingly evident, impacting the very nature of the interactions between users and online services and content. Representing a new form of intermediation, these agents are transforming the way users access and interact with online resources, while also reshaping the distribution of informational power, economic value, and access in the digital ecosystem. This paper analyses the evolution of intermediation mechanisms on the internet through a critical, non-systematic review guided by Critical Systems Heuristics. We trace the progressive layering of intermediation mechanisms to online content/services from the technical infrastructures, and we argue that AI agents constitute a structurally distinct intermediation layer that introduces new dynamics of power, opacity, value distribution, and access to the digital ecosystem, with significant implications for users, online services, and the broader society. We are currently crossing a potentially significant evolutionary shift in the Internet, where AI agent intermediation could progressively exert a substantial influence on how information is exchanged. This paper offers a framework with which to discuss the implications of this potential evolution and to guide future research, public discussion, and policy interventions in this area.
In IoT health monitoring using electrocardiograms (ECGs), the surge in data transmission volume poses a significant challenge. This study utilized PhysioNet ECG data to compare the transmission volumes of raw ECG signals, R-R intervals (RRIs), and HRV metrics (SDNN, RMSSD, and LF/HF), thereby evaluating the effectiveness of communication optimization. The results demonstrated that transmitting RRI data and HRV metrics reduced data volume by approximately 99% and over 99.9%, respectively, compared to transmitting raw ECG data. These results suggest that the proposed approach could contribute to improved energy efficiency and reduced transmission latency in wearable devices, supporting its potential feasibility for bandwidth-constrained IoMT deployments.
Public electric vehicle (EV) charging increasingly relies on internet-connected platforms for metering, billing, roaming, and settlement. However, final billing records and charge detail records often provide limited evidence that a session result can be independently linked to the ordered metering data from which it was derived. This paper presents a reproducible blockchain-anchored Proof-of-Charge platform for generating tamper-evident and auditable EV charging receipts. The platform converts charging-session data into canonical receipts, computes cryptographic commitments over receipt content and ordered meter values, aggregates receipt hashes using a temporally ordered and domain-separated Merkle profile, and anchors compact batch commitments in a smart contract while keeping detailed records off-chain. A working prototype implements cross-language canonicalization checks, membership-proof generation, structured storage, batch anchoring, verification services, synthetic workload generation, dataset export, and local blockchain deployment. Across 50 measured runs covering 10 to 1000 receipts, all count reconciliations, batch-root checks, and on-chain comparisons passed. Mean receipt-pipeline latency ranged from 7.152 to 7.853 ms per receipt, with throughput from 127.54 to 141.23 receipts/s. A focused 1000-leaf proof sample produced a 2064-byte proof with ten sibling hashes. The results show that the platform can generate, anchor, and verify auditable EV charging receipts with reproducible performance while keeping detailed charging data off-chain. The proposed architecture provides a practical digital trust layer for internet-enabled EV charging and V2G-ready settlement workflows.
Speech emotion recognition (SER) is a fundamental task in affective computing; however, traditional unimodal approaches often struggle to capture the complex emotional cues present in spontaneous conversational speech. Bimodal frameworks that integrate acoustic and textual information have therefore emerged to provide complementary semantic and acoustic representations. This study proposes a bimodal SER framework based on a hybrid convolutional neural network–long short-term memory (CNN–LSTM) architecture. Using the Multimodal EmotionLines Dataset (MELD), the framework combines temporal acoustic features, statistical acoustic features, and predicted textual sentiment. Experimental results indicate that the proposed model achieves reliable recognition of majority emotion classes but exhibits limited performance on underrepresented minority classes due to severe class imbalance. To better understand the contribution of each modality, feature sufficiency and feature necessity analyses were conducted. Furthermore, an evaluation of alternative fusion strategies showed that the expressive attention networks did not provide meaningful performance improvements over simple feature concatenation. These findings suggest that class imbalance, rather than fusion complexity, remains the primary limitation in conversational SER, highlighting the importance of addressing data imbalance before pursuing more sophisticated multimodal architectures.
Human Activity Recognition (HAR) based on Channel State Information (CSI) has attracted considerable attention as a privacy-preserving sensing paradigm. However, CSI-based HAR faces several challenges, including environmental noise, long-range temporal dependencies, and the anisotropic structure of CSI tensors. To address these challenges, this paper presents Wi-CSNet, a lightweight CSI-oriented framework that integrates Discrete Wavelet Transform (DWT) preprocessing, asymmetric-stride convolutions, and a Cross-Scanning State Space Duality (CS-SSD) block derived from Mamba2. DWT preprocessing compresses temporal signals while preserving motion-related trends and reducing input dimensionality. Asymmetric-stride convolutions balance feature scales across heterogeneous CSI dimensions, while the lightweight CS-SSD module captures global dependencies with only a 0.63% parameter overhead. Extensive experiments demonstrate that Wi-CSNet achieves accuracies of 97.81% on HHI, 99.92% on UT-HAR, and 100% on NTU-HAR. These results confirm the effectiveness and robustness of Wi-CSNet for fine-grained CSI-based activity recognition in complex environments.
