Reliable deployment of analytical decoders and artificial intelligence (AI) models in resource-constrained edge AI systems requires verification beyond task accuracy. Model conversion, quantization, runtime integration, backend-specific preprocessing, and target-side execution can alter output vectors or class decisions even when a backend remains executable. This paper presents a host-to-MCU deployment-fidelity study of 11 analytical, floating-point neural, and quantized neural steady-state visually evoked potential (SSVEP) backend realizations on an ESP32-S3 under strict leave-one-subject-out replay of the 12JFPM dataset. The evaluation combines trial-aligned host–device top-1 agreement and within-backend scale-normalized output-vector deviation with stage-resolved checkpoint-to-exported-host and exported-host-to-device analysis, preprocessing-inclusive latency, firmware footprint, and runtime-memory profiling. All floating-point neural backends preserved top-1 decisions across both neural deployment comparisons. Int8 disagreement was backend dependent: CCNN-int8 showed greater disagreement during export-to-device realization, whereas EEGNet-int8 showed greater disagreement at the checkpoint-to-exported-host boundary. Analytical backends achieved high decision preservation but varied substantially in latency and persistent resource demand. Illustrative hard-constraint policies selected different realizations as fidelity, latency, precision, and memory requirements changed. These results show that successful execution and task accuracy alone do not establish faithful embedded realization. Behavioral preservation, target-side trial-processing cost, and measured resource demand should be evaluated jointly when selecting heterogeneous backends for resource-constrained computing.
Software-Defined Networking (SDN) has become a widely used approach for centralized network control in cloud environments, yet most deployments still operate reactively. This paper presents a proactive control framework in which OpenStack Nova lifecycle events are used as direct network management signals. Events delivered via the oslo.messaging bus are evaluated by a rule-based decision engine, which applies Neutron QoS policies without waiting for telemetry-based degradation triggers. Dynamic port resolution ensures that policies are applied to the correct VM port at runtime. Framework was built and evaluated on logically multi-node OpenStack 2025.1 environment. For 27 measured VM creation events, QoS configuration averaged 6,981 ms from Nova notification receipt to Neutron QoS API completion. Rule evaluation added almost no processing overhead ($< 0.03$ ms), while QoS API calls took 568 ms on average. Main delay was associated with Neutron port availability after the create.start notification, with port resolution taking 6,412 ms on average. Results show Nova lifecycle events can configure QoS policies in OpenStack without changing existing infrastructure. We also compared the event-driven approach with periodic polling at intervals of 5, 10, 30, and 60 s. Results varied depending on polling interval and when Neutron port became available. At 5 s, both approaches took approximately 11.8 s. However, the event-driven approach was faster by 1.29 s, 6.75 s, and 19.84 s at polling intervals of 10, 30, and 60 s, respectively. For resize events, average latency was 730 ms, mainly because Neutron port was already available. This indicates that port availability, rather than controller processing, accounts for most of the delay observed during VM creation.
Immersive metaverse environments must reconcile real-time rendering, spatial privacy, and verifiable identity. Centralized providers concentrate biometric telemetry into single points of failure; soulbound tokens expose it on public ledgers to motion-based de-anonymization. This work presents a Groth16-based zero-knowledge identity architecture separating two timing domains: proof verification runs locally inside the render loop before session access is granted, and nullifier settlement follows asynchronously for replay protection. The chain is not in the critical path of the access decision. On consumer-grade hardware, the 2,163-constraint circuit verifies in 8.41 ms on average (N=100), within the 20 ms motion-to-photon budget. A client-side bounded noise mechanism, evaluated on 200 synthetic subjects calibrated to published VR motion statistics, produces full commitment churn at every nonzero noise level and cuts k-NN re-identification from 99.85% to 73.60%. Replayed on 276 real BOXRR-23 users, churn replicates exactly while baseline re-identification is an order of magnitude lower (9.02%), falling to 1.99% under noise: per-feature Gaussian calibration reproduces marginal distributions but overstates inter-subject separability, so the synthetic figures measure an idealised population, not the real threat. The benchmarked circuit binds each session to one enrolled key but does not prove registry membership in-circuit, so Sybil resistance rests on off-chain issuance; the Merkle-inclusion extension that would close this gap is specified and left to future benchmarking. The contributions are a STRIDE/LINDDUN threat model for the ternary metaverse identity pipeline, an MTP-constrained Groth16 benchmark, and a session-commitment churn analysis validated on synthetic and real telemetry.
As semiconductor devices progress into the Ultra-Deep Sub Micron (UDSM) technologies, Network-on-Chip (NoC) interconnections face severe vulnerability to transient faults caused by noise-induced errors, posing a significant threat to data integrity and system reliability. These soft errors represent one of the major hurdles in the successful implementation of many core processor systems in applications that require rigorous safety standards, such as automotive systems, aerospace systems, and medical devices. Although Error Correction Codes are a widely accepted counteractive strategy, currently known solutions come with prohibitively high overheads in terms of average latency, power, and area or are found to be unable to cope with the most common multi-bit random error patterns. The Improved Hamming-based Multiple Transient Error Correction (IHMTEC) technique is based on a unique non-contiguous bit mapping and a novel iterative decoding process. Upon being evaluated using 45nm technology with a standard 32-bit flit under an independent random bit flip error model, our technique establishes comprehensive correction of all single, double, and triple bit errors, which have been verified using exhaustive pattern testing. Furthermore, using extensive simulations, it has shown successful correction of an average of 99.5% with a 95% confidence interval of +/- 0.02% of all possible 4-bit and 5-bit random error patterns within a 32-bit flit. This is validated using Monte Carlo bit-flip injection under an independent and identically distributed Bernoulli bit-flip model with N=$10^{9}$ trials. This work offers enhanced reliability with competitive area, power, and average latency metrics, supporting sustainable, resilient, and energy-efficient semiconductor infrastructure. The detailed analysis indicates that the IHMTEC technique improves the reliability of nanoscale Network-on-Chip interconnects, enabling scalable on-chip communication for next-generation reliable computing infrastructures.
Finding nutrient deficiencies in crop leaves is important for precision agriculture because it enables timely intervention and improved fertilizer management. This study presents NutriSafe Net (NSNet), a compact lightweight deep learning framework for the classification of nutrient deficiencies in crop leaves. NSNet integrates a Multi-Scale Convolutional Spatial Attention Fusion (MCSAF) module with a Support Vector Machine (SVM) classifier to enhance feature extraction and classification performance while maintaining a small computational footprint. The MCSAF module employs multi-branch convolutions ($1 \times 1$, $3 \times 3$, and depthwise $5 \times 5$), augmented with the Convolutional Block Attention Module (CBAM), to capture complementary features across multiple spatial scales. Instead of a conventional softmax layer, NSNet uses an SVM classifier to improve class separation for visually overlapping nutrient-deficiency categories. The model was evaluated on a five-class nutrient-deficiency dataset using five-fold cross-validation and achieved an accuracy of 99.21%. NSNet outperformed well-known lightweight and transformer-based models such as MobileNetV3 (97.69%), EfficientNet-B0 (98.11%), MobileViT-S (98.11%), Swin Transformer (97.9%), and ViT (97.8%). In addition, NSNet has a compact footprint of only 0.08 million parameters and 435 MFLOPs. To assess its edge feasibility, NSNet was deployed on a Raspberry Pi 5 equipped with a DKP industrial camera, achieving real-time inference at 6.7 FPS with a latency of 149 ms while accurately identifying nutrient deficiencies at a practical working distance. These results demonstrate that NSNet provides a favorable balance between accuracy, efficiency, and model compactness, making it a promising framework for resource-constrained smart farming applications.
