Effective detection of network anomalies is crucial when it comes to security of computer networks, but traditional methods tend to fail when used to address a wide range of traffic and dynamically changing conditions. This paper provides a systematic review of eight ensemble algorithms, including Random Forest, Extra Trees, Bagging, AdaBoost, Gradient Boosting, HistGradientBoosting, Stacking and Voting, on a dataset of 4,998 samples where 35 features were statistics of network traffic. The data underwent preprocessing based on cleaning, normalization and encoding, and we evaluated the models based on stratified 10-fold cross-validation which used accuracy, precision, recall, F1-score and AUC. Our findings reveal that Stacking with a meta-ensemble produced the best results of 98.90 percent and Voting closely followed with 98.85 percent. Random Forest and Extra Trees also showed strong results of more than 98 percent, whereas the weakest result of 78 percent was obtained in Gradient Boosting, which is sensitive to the configuration of the data. These results offer solid empirical support that ensemble architectures, specifically stacking and voting, offer highly precise, scalable, and practical solutions to state of the art intrusion detection systems.

The intensive growth of Internet of Things (IoT) devices has exponentially increased cyber-attack surfaces, and the resource-constrained nature of IoT-based devices strongly restricts the ability to deploy traditional deep-intrusion detection systems (IDS). In this paper, a lightweight hybrid IDS is proposed, which consists of a light convolutional neural network (CNN) combined with bidirectional long short-term memory (BiLSTM) and Incremental Principal Component Analysis (IPCA) to perform online dimensionality reduction on features. The offered method is considered in detail using both real-world datasets of IoT intrusion, namely CICIoT2023 (large-scale lab-generated IoT attacks with 33 attack types) and IoT-23 (realistic long-duration malware scenarios on commercial IoT devices). The model achieves a detection accuracy of 98.23% and a recall of 98.6% on CICIoT2023. On the IoT-23 dataset, it yields a detection accuracy of 97.15% and a recall of 97.1%, indicating that it can generalize more strongly across different distributions of IoT traffic. The method reduces model size by 60% and inference time by 65% compared to full-feature deep baselines, and achieves better accuracy than the current state-of-the-art lightweight methods. The findings indicate the feasibility of the method in effective and real-time IoT intrusion detection using a limited edge infrastructure.

The basic objection to smart contracts introduction in areas where data secrecy is essential, e.g. finance and healthcare, is the transparency nature of blockchain technology. Although zero-knowledge proofs (ZKPs) provide a promising solution to verifiable privacy, they have high computational and financial complexity levels, which make them unfeasible to use universally. This paper presents an adaptive multi-layer encryption structure to determine a solution to this limitation. These dynamically chosen and constructed cryptographic primitives (such as symmetric encryption, commitment, and zk-SNARKs) are informed by an intelligent policy engine and are selected and constructed according to the sensitivity of the transaction and the real time network conditions. This approach is no longer a traditional balance between performance and security because it uses a context-aware structure that is no longer a simple one-size-fits-all approach to privacy. We deploy a prototype to an Ethereum testnet and show that our framework supports a 38.7% and 86.9% reduction in computation overhead and gas expenditure in performing low-sensitivity transactions, respectively, over mandatory ZKP baselines, as well as strong privacy guarantees in sensitive operations. The findings affirm that adaptive, multi-layer cryptography is a sensible direction towards effective and realistic confidential smart contracts to be used in large scale application. Future research will be the incorporation of post-quantum cryptographic tools and how they can be customized to decentralized processes to achieve long-term security sustainability.

The rapid proliferation of artificial intelligence in mental health applications, particularly in digital journaling and emotion tracking, has accumulated significant privacy concerns regarding the centralization of sensitive user data. Although Federated Learning offers a decentralized alternative by training models locally on edge devices, existing frameworks predominantly rely on the assumption of Independent and Identically Distributed unimodal data. This approach fails to address the inherent diversity of real-world user interactions where client inputs vary dynamically between text-only entries and multimodal content. To bridge this gap, we present MobileFedFusion, a resource-efficient and privacy-preserving Federated Learning architecture designed for heterogeneous edge environments. We propose a novel Modality Masking mechanism that enables a unified global model to aggregate gradients from diverse clients, seamlessly integrating text-only and multimodal contributors without architectural fragmentation. The system leverages a lightweight fusion of MobileBERT and MobileViT that is specifically engineered to operate within the computational constraints of mobile hardware. Experimental validation on a non-IID partition of the Reddit GoEmotions and Memotion 3.0 datasets demonstrates that our approach achieves a Global Macro F1 score of 0.68. The results indicate that the model effectively converges to state-of-the-art performance while ensuring that sensitive personal data never leaves the local device.

Session-Based recommendation models such as item Based Collaborative Filtering (Item CF), Session MF, and Bayesian Personalized Ranking (BPR), work but do not catch sequential patterns. Systems that use recurring patterns such as RNN and GRU4Rec do better with order but have trouble with long-distance item links and take a long time to train. Neural Collaborative Filtering (NCF) offers deeper User-Item connections but misses session context, while Graph Neural Network methods can show complex structures but need heavy calculations and can smooth things out too much. To fix these problems, we suggest the Lightweight Graph Attention (LGAT) Model. LGAT shows each click sequence as a directed session graph with items as nodes and moves between them as edges; it then uses a single Self Attention layer to figure out how important each edge is. A simple attention-pooling method then brings node embeddings together into a small session vector. The model predicts the next item using dot product scoring and trains from start to finish with Cross entropy loss. In three tests (DIGINETICA, Retail Rocket, Yoo Choose 1/64), LGAM did better than six other models and got the best results for HR @10, NDCG @10 and MRR @10 while cutting training time by 20-30%. LGAT’s novelty lies in its combination of graph structure and lightweight attention, yielding both high accuracy and real-time efficiency.

Software Fault Prediction (SFP) plays a very crucial role in improving software reliability by facilitating the early detection of modules prone to defects. Nevertheless, ongoing issues like extreme imbalance in classes and unstable performance of the classifiers on the heterogeneous datasets deter the efficiency of current methods. To address these problems, in this paper, a stability-conscious meta-ensemble learning architecture is proposed combining adaptive sampling with meta-level classifier fusion. Contrary to traditional ensemble-based approaches that rely on resampling and fixed combinations of models, the presented architecture dynamically chooses the appropriate sampling techniques to rely on the properties of the data and trains the best combination of classifiers with the help of a meta-learner. Wide experiments performed on benchmark datasets of PROMISE, NASA, AEEEM, ReLink, and SoftLab indicate that there is a consistent improvement in performance compared to baseline ensemble models with better AUC, MCC, and G-mean. Moreover, the experiments of the cross-project fault prediction prove high generalization and low deterioration of performance. The statistical significance tests such as Wilcoxon Signed-Rank Test, Cliff- Delta, and Nemenyi post-hoc tests confirm the strength of the suggested method. In general, the framework offers a practical and generalizable method of resolving the issues of class imbalance and performance instability in the real-world software fault prediction.