Device heterogeneity in federated fine-tuning of vision-language models (VLMs) leads clients to adopt different LoRA ranks, making conventional homogeneous aggregation difficult to apply. Existing approaches either impose a common rank across clients, limiting resource adaptivity, or reconstruct full-dimensional matrices update with singular value decomposition (SVD), resulting in substantial server-side computational overhead. We propose FedHRQS, a resource-aware federated LoRA framework that supports client-specific rank selection and heterogeneous-rank aggregation through sample-weighted factor concatenation. FedHRQS applies dual thin QR decomposition to the concatenated LoRA factors and performs SVD on the resulting compact core matrix. We show that this design yields the same truncated-SVD solution as directly decomposing the full aggregated update matrix, while substantially reducing server-side complexity by restricting SVD to a compact matrix whose dimension is determined by the total LoRA rank rather than the full model dimension. We further derive a generalization bound characterizing the effects of client-specific rank truncation and effective model dimensionality. Experiments with LLaVA-1.5-7B on LLaVA-ReCap-118K show that FedHRQS improves Google BLEU by up to 40.2 × –7586 × speedups in server-side aggregation, which demonstrates that FedHRQS effectively accommodates device heterogeneity while enabling accurate and computationally efficient federated fine-tuning of VLMs.
The proliferation of Internet of Things (IoT) and Internet of Medical Things (IoMT) devices has dramatically escalated the risk of cyberattacks, necessitating robust and highly responsive Intrusion Detection Systems (IDS). However, existing solutions frequently struggle with severe class imbalance and high-dimensional data, leading to a suboptimal trade-off between detection accuracy and computational efficiency. To address these challenges, this paper introduces a novel hierarchical machine learning-based IDS optimized for real-time multiclass attack detection. We construct a comprehensive unified dataset by integrating the recent CICIoT2023 and CICIoMT2024 datasets. Our framework first employs a Genetic Algorithm (GA) for feature selection, extracting a lightweight subset of critical features to minimize computational overhead. The core classification engine utilizes a two-stage hierarchical Random Forest architecture coupled with selective SMOTE oversampling. The first stage identifies major attack categories while grouping minority instances into a meta-class, which are then precisely classified by the second stage. Experimental evaluations demonstrate that the proposed hierarchical framework establishes an optimal accuracy-efficiency trade-off and improves the classification performance on minority attack classes compared to traditional flat classifiers. While simple data undersampling methods achieve fast runtimes at the cost of severe performance degradation, our hierarchical approach preserves exceptional detection capabilities and operates faster than conventional models trained on full datasets. The framework achieves robust F1-scores of 90
In Wireless Sensor Networks (WSNs), secure and energy-efficient data aggregation has been challenging due to limited energy supply, dynamic nature of the network, and presence of malicious nodes. In this paper, a Binarized Simplicial Convolutional Neural Networks for Secure and Energy-Efficient Data Aggregation in WSN (CPGIMC-SEIDA-WSN) is proposed. Initially, WSN nodes are collected from real-time environmental data like humidity, temperature and motion. The Consensus Partition Guided Incomplete Multi-View Clustering (CPGIMC) used to format the clusters and selects the optimal CH through Parameter Adaptive Manta Ray Foraging Optimization (PAMRFO). Malicious nodes are identified and isolated using novel Binarized Simplicial Convolutional Neural Networks (BSCNN), which analyze the topological relationships among sensor nodes and detect abnormal communication patterns to ensure data integrity. The Superb Fairy-Wren Optimization Algorithm (SFOA) also optimizes the path of the data transmission through three parameters like trust, connectivity and quality of service (QoS). The simulation proves that the proposed approach has the capability to provide 23.55
Vehicle Ad-hoc Networks (VANETs) are fundamental to achieving zero-fatality intelligent transportation, yet their open nature exposes them to severe security threats, particularly blackhole attacks that disrupt routing and compromise safety. Existing countermeasures force a trade-off between security and efficiency: cryptographic methods impose excessive overhead, while machine learning detectors lack real-time adaptability. To bridge this gap, this paper proposes AC-NB, a closed-loop framework that integrates an enhanced Ant Colony Optimization (ACO) routing protocol with a lightweight Naïve Bayes classifier. In AC-NB, a continuous feedback cycle is established: the Bayesian detector computes node trust values from packet drop rate, forward rate, and RSSI mismatch, and these trust values directly update the ACO heuristic—which jointly considers link security, link quality, distance, and node trust—steering route selection away from malicious nodes. Extensive highway simulations under 20
Network intrusion detection systems (IDS) are still a staple of the contemporary cybersecurity infrastructure, but the traditional methods based on fixed signatures and theoretical rule sets are still not able to overcome the high false positive rates, slow detection and a general inability to follow the fast changing tactics used by attacks. These are especially acute in settings where encrypted traffic streams, heterogeneous IoT devices and advanced adversarial interactions that strategically bypass signature-based defenses are present. In this paper, the author presents the Maintainable Adaptive AI (MAI) model, a new two-stage ensemble architecture, which combines gradient boosting (XGBoost) with a lightweight feedforward neural network to provide real-time continuously adaptive intrusion detection. The MAI model also uses dynamic feature selection of the lexical, host-based, traffic-based, and statistical attributes in addition to an incremental learning process directed by the KL-divergence-based drift detection. In contrast to earlier, ad hoc proposals of adaptive IDS, the MAI framework is specifically maintainable by a selective retraining of components that only increases the weight of the neural networks when in standard operation and initiates a complete ensemble recalibration when distribution drift is beyond predefined thresholds. Thorough testing using four different sets of data, which are the Malicious URLs Dataset (651,191 entries), ISCX VPN-NonVPN Traffic Dataset (200,000 samples), TON IoT Dataset (400, 000 samples), and a newly curated Custom Scraped Dataset (cumulative of approximately 200,000 real-world attack logs across open-source threat intelligence feeds) prove precision of 93–96, recall of 92–95 and overall Formal complexity analysis The computational tractability of the model is defined and inference latencies of 1.8 ms per sample with batched computation on a GPU, which can be deployed at network links with throughput as high as 4.3 Gbps. Initial adversarial robustness evaluation shows that the model still achieves more than 83
