Prior to the advent of fault tolerant quantum computers, the ability to execute post quantum cryptographic (PQC) algorithms on noisy intermediate scale quantum (NISQ) devices is crucial to evaluate their security and performance. However, time-varying errors in the gates and measurements not mitigated by static means can compromise cryptographic security, resulting in unacceptably high decapsulation failure rates. This paper presents an adaptive error mitigation framework FF that involves an online noise estimation algorithm, a reinforcement learning agent and a library of mitigation actions such as readout error inversion, probabilistic error cancellation, Clifford assisted transformation, dynamical decoupling, zero noise extrapolation, and qubit remapping. The framework is dynamic and selects optimal actions based on real-time estimates of noise to minimize the total variation distance (TVD) subject to gate overhead and latency constraints. We benchmark FF across three NISQ devices: IBM Brisbane, Rigetti Aspen M 3, and IonQ harmony on the circuits Kyber 512, Dilithium 2 and SPHINCS+ 128f. The adaptive framework improves the TVD by 44-51% when compared to the best static method and 78-83% when compared to no mitigation. Most importantly, only the adaptive framework guarantees Kyber 512's decapsulation failure probability to be below the 2^-128 security threshold on all devices, whereas static methods are found to be above the bound because of the unmodeled noise drift. Gate overhead is kept within 20% and classical latency under 50ms per slice. These results demonstrate that adaptive, learning-driven error mitigation is not merely beneficial, but necessary for provably secure PQC execution on NISQ hardware. The framework and benchmark suite are released open-source to accelerate reproducible research.

The rapid growth in the number of deployed Internet of Things (IoT) networks has made these devices more vulnerable to emerging cyberattacks, such as distributed denial-of-service, botnet, spoofing, reconnaissance, brute-force, malware, and data injection attacks. Centralized intrusion detection frameworks typically require the collection of traffic data in a central server to perform analysis and detection tasks. This process may involve potential privacy risks, large communication overheads, delayed detections, and scalability issues owing to the heterogeneous clients in IoT networks. To mitigate these problems, this study proposes a Dependability-Aware Adaptive Federated Learning-Based Intrusion Detection System (ADFed-IoTIDS) for detecting evolving cyberattacks in IoT networks. In this framework, several IoT clients or gateway nodes learn local intrusion detection models based on private traffic data, whereas only model updates are transmitted to the federated aggregation server. An adaptive aggregation method is introduced to enhance the performance of global learning models by considering aspects such as local learning performance, non-independent identically distributed (non-IID) data distribution, client reliability, local loss, validation scores, and attack diversity. The proposed framework was tested using benchmark IoT intrusion detection datasets, such as CICIoT2023 and Edge-IIoTset. The evaluation criteria included classification metrics such as accuracy, precision, recall, F1-score, and dependability metrics such as false alarm rate, detection delay, communication overhead, scalability, reliability and stability. The anticipated outcomes indicate that the ADFed-IoTIDS framework will be a better intrusion detection framework than centralized machine learning models, local learning models, FedAvg, and FedProx frameworks.

The rapid proliferation of Internet of Things (IoT) devices has significantly increased the cybersecurity attack surface, making IoT environments highly vulnerable to sophisticated botnet-based attacks such as Mirai and Gafgyt (BASHLITE). Although conventional intrusion detection systems (IDS) have demonstrated promising performance, they often depend on large-scale labeled datasets and primarily focus on binary attack detection, limiting their applicability in realistic IoT environments characterized by scarce attack data and evolving threats. To address these challenges, this study proposes a hybrid quantum-classical intrusion detection framework, termed KPCA-VQC-XGBoost, for sample-efficient multiclass IoT botnet detection. The proposed approach integrates Kernel Principal Component Analysis (KPCA) for nonlinear feature reduction, a 10-qubit Variational Quantum Circuit (VQC) for quantum-enhanced feature extraction, and XGBoost for robust multiclass classification. The framework is evaluated on the N-BaIoT dataset to distinguish Benign, Gafgyt, and Mirai traffic classes using only 2,400 stratified training samples. Experimental results demonstrate that the proposed framework achieves 95.70% classification accuracy, a macro F1-score of 0.960, and Area Under Curve (AUC) values greater than 0.994 for all classes, indicating strong generalization under constrained training conditions. Comparative analysis further reveals that the proposed framework maintains competitive performance while substantially reducing training data requirements relative to existing quantum intrusion detection approaches. The findings highlight the potential of hybrid quantum-classical learning architectures for developing lightweight, scalable, and data-efficient cybersecurity solutions for next-generation IoT systems.

A computation-aware tiered eviction system is proposed for managing key-value (KV) caches during large-language model inference. A major difficulty lies in the fact that standard eviction policies, such as the least-recent-used algorithm, do not consider the differing computational expenses of recalculating distinct KV blocks after their transfer to slower memory. Our method introduces a Computation-Aware Eviction Controller that assigns a score to each KV block according to its attention variance, depth of dependency within the transformer, and temporal locality. A secondary cache backed by an SSD is used for lower-scoring blocks to keep them out of high-bandwidth memory, whereas high-scoring blocks are kept in high-bandwidth memory. A lightweight runtime profiler updates these scores with minimal overhead. Furthermore, a predictive prefetch mechanism retrieves blocks from the SSD that are likely to be needed before they become necessary. The system achieved a 1.8x improvement in sustained throughput for sequences exceeding 32K tokens compared to a baseline LRU policy, without degrading generation quality. This work shows that embedding computational expense into cache handling is a feasible and efficient method to resolve memory constraints in extended-context inference.

Energy-efficient virtual machine (VM) consolidation is an important part of cloud data center operational overhead reduction. This paper investigates the use of a hybrid consolidation framework based on the Modified Best Fit Decreasing (MBFD) algorithm for initial VM placement and a Q-learning based decision model for dynamic VM reallocation. A combination of the Q-learning and the dual-threshold policy is used to select migration actions based on system states, which include CPU load, power characteristics, and parameters that are related to SLA. A thorough experimental analysis is performed in the MATLAB environment in various workload situations, where the proposed technique is compared to well-established algorithms such as E-ABC, OGrA+OFr and DA-MBFD. The performance trends are analyzed in terms of energy consumption, violations of SLA, number of migrations and utilization of hosts. The results offer a detailed comparison of the behavior of Q-learning-based consolidation as compared to heuristic and metaheuristic strategies, which provides insights into energy-aware resource management in the cloud environment.

Modern industrial cyber-physical systems (CPS) rely on dense sensor networks whose telemetry is routinely corrupted by measurement noise, transmission jitter, sensor drift, and occasionally adversarial perturbation. Such corruption degrades not only the raw accuracy of learning-based fault detectors but also the broader performability of the monitoring system that it’s intended to deliver correct, timely, and available detections under degraded operating conditions. This paper introduces PerforNet, a noise-resilient, attention-guided contrastive learning framework that jointly optimizes detection accuracy and quantified system performability. PerforNet combines (i) a dual-branch channel-temporal attention encoder that learns to down-weight historically unreliable sensor channels rather than relying solely on instantaneous signal salience, and (ii) a performability-weighted contrastive objective that discounts the contribution of negative pairs whose apparent dissimilarity is attributable to noise rather than true semantic (fault) difference. We further propose a Composite Performability Index (CPI) that unifies detection F1-score, false-alarm rate, and operational availability into a single interpretable benchmarking score, bridging machine-learning evaluation with classical reliability engineering. Experiments across three CPS benchmarks under controlled sensor-noise injection show that the proposed framework improves detection F1-score over strong contrastive and reconstruction-based baselines while substantially reducing false-alarm frequency, with ablations attributing the majority of the robustness gain to the performability-weighting term rather than the attention module alone. The framework offers a practical and reproducible route toward performability-aware design of intelligent monitoring systems for safety-critical infrastructure.