International Journal of Human Computations and Intelligence

Digital Forensics for Detecting and Investigating Cyber Malicious Activities

Marathi Muni Babu — 2026-02-17

Cybersecurity is an urgent concern in this age of rapid expansion of digital infrastructures, especially due to insider threats. These are sophisticated threats where traditional signature-based detection methods have proven much less effective, since these attacks are by people who have legitimate access to sensitive data. In this paper, several ML models, namely Logistic Regression, Random Forest, Support Vector Machine, Decision Trees, and XGBoost, have been experimented with for detecting insider threats in cybersecurity. XGBoost proved to be the best among the compared ML models, with an accuracy of 94.2%. However, the proposed model CNN outperformed all other algorithms and achieved 95.0% accuracy along with the highest precision and F1-score. This confirms that deep learning techniques are much better at capturing complex patterns in cyber activities than ML techniques. While the proposed CNN resulted in excellent performance, several challenges remain to be explored, such as the problem of class imbalance, anomaly detection in real time, and explanation of anomalies. This paper presents a proposal that the integration of advanced machine learning and deep learning models is crucial for improvement in scalable, real-time, and accurate cybersecurity solutions.

Reinventing Authentication through User-Centric Two-Factor Security and Personalized Image Verification

J Sivarani — 2026-02-18

In the wake of the rising digital world, traditional username/password-based authentications are facing an increased risk of being compromised by various cyber attacks, including phishing, brute force attacks, credential stuffing, and artificial intelligence-based impersonation attacks. While traditional two-factor authentications provide better security than traditional username/password-based authentications, these methods often bring their own set of problems, including usability issues. To overcome the limitations of traditional two-factor authentication, this paper proposes a two-factor authentication system based on traditional username/password-based authentication and image/keyword-based authentication. The system utilizes the capabilities of human visual memory to provide better security with minimal cognitive overhead for the users. The system utilizes secure cryptographic techniques to ensure the security of the system, keyword-based image verification to prevent replay attacks, and device-based restriction techniques to prevent brute force attacks. The system is developed based on the client-server model using the Django framework. The experimental results confirm that the proposed authentication system achieves an optimal balance between usability and security, making it a feasible solution for web application security.

Predicting Cybersecurity Risk Through Fuzzy and Adaptive Neuro-Fuzzy Systems

P Bhanu Prakash — 2026-02-24

As modern networks grow in intricacy, the cybersecurity risks involved are becoming increasingly difficult to foresee and handle. The dynamic, uncertain, and nonlinear nature of these risks cannot be coped with using traditional rule-based and static risk assessment methods. We suggest a Hybrid Cybersecurity Risk Prediction Model through the use of both Fuzzy Inference System (FIS) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for appropriate prediction as well as risk assessment. This is because FIS tends to offer an interpretable approach towards risk assessment using expert-defined rules, thereby making predictions more comprehensive. On the other hand, the incorporation of the neural network in ANFIS makes predictions and risk assessment accurate. Our model extracts the critical cybersecurity parameters, comprising Authentication Strength, Traffic Anomaly Levels, and Network Load Conditions, from the UNSW-NB15 dataset. Experimental results show that the hybrid FIS-ANFIS model gives a high prediction accuracy, robustness, and adaptiveness over traditional machine-learning models, including. We present the effectiveness of ANFIS in coping with uncertain and evolving cybersecurity threats as a promising tool for real-time cybersecurity risk management. These results strongly substantiate the proposed model significantly outperforms conventional models in reliably and precisely making predictions about cybersecurity risks within complex network environments.

A Robust IDS-Driven Approach to Secure Industrial Control Systems from Emerging Cyber-Physical Attacks

D Nagabhushanam — 2026-02-24

CPS are critical for the operation of many infrastructures, such as ICS, energy, and transportation. However, increasingly interconnecting cyber with physical elements opens the floor for complicated cyber-physical attacks targeting severe operational disruption. This paper proposes an efficient approach, basically IDS-driven, for securing ICS against such advanced threats. A lightweight hybrid IDS is proposed for the efficient detection of known attacks as well as unknown zero-day attacks using supervised and unsupervised training models, respectively. The suggested IDS incorporates the LightGBM and Isolation Forest models for signature-based detection and anomaly-based detection, respectively. This will ensure that the threat detection is comprehensive in dynamic industrial environments with timely identification and response against both traditional and evolving cyber-physical threats. In our experimental results, the proposed H-IDS was evaluated using the NSL-KDD dataset and outperformed the traditional approaches in terms of accuracy, precision, and recall. Moreover, it will bring benefits like reduced false positive rates, speed, and resiliency. Some possible improvements that can be made in the future are also discussed in the paper, such as incorporating edge computing with federated learning, which can enhance capabilities for the proposed system in real-time environments.

A Unified Architecture for Real-Time Analytics Using Microsoft Fabric OneLake

Narendra Mangala — 2026-02-24

In this study, in the day and age of digital disruption, there is a substantial task at hand regarding the volume and speed of data associated with the Internet of Things (IoT), financial transactions, and logs. A complex architecture with complex ETL (Extract, Transform, Load) and disconnected storage schemes typically causes substantial data latency, hindering real-time decision support. To overcome these challenges, this paper presents an integrated architecture that leverages Microsoft Fabric OneLake, a common data lake infrastructure that requires no physical data transfer via shortcuts, serving as a starting point for fast, real-time analytics. The backbone of this architecture consists of an intelligent analytics layer incorporating a Transformer Autoencoder. Contrary to traditional approaches that use Linear Regression (LR), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU), which rely on self-attention in Transformers to directly compute the temporal importance of all timestamps, this model achieves revolutionary accuracy in characterizing seasonality and long-term dependencies in multivariable time series data. The algorithm includes an unsupervised learning-based rebuilding strategy that identifies anomalies when corresponding activities in a system fall above or below a data-driven threshold for normal operations. Experiments performed on a smart grid real-time load data test set reveal that the proposed strategy using a Transformer Autoencoder attains maximum accuracy, precision, and F1 scores of 96.0%. This result clearly beats all comparison algorithms, namely Linear Regression (88.5), Isolation Forest (90.8), LSTM (92.3), and GRU (93.4). Additionally, this strategy is integrated with Microsoft Fabric Real-Time Intelligence workload.

An Efficient Machine Learning Framework for Behavioral Insider Threat Detection with Comparative Analysis of Ensemble Methods.

J Sivarani — 2026-03-02

The insider threats, where insiders with access and authorization misuse it, pose a major challenge to the current prevalent cybersecurity systems. With respect to the insider threats, it can be noted that they have access and authorization, and hence, detection is quite complex and challenging, unlike other attacks by external users. The current analysis focuses on the detection of these insider threats using quite effective and efficient machine learning classifiers, particularly the Decision Tree, Random Forest, and XGBoost classifiers. These classifiers have been chosen because they can handle large volumes of data and can easily cater to the detection of the various insider threats by identifying the pattern or anomaly with respect to the user access. The decision tree is quite efficient and can be easily interpreted, and hence used, whereas the random forest classifier combines the results and predictions made by each and every decision tree, thus giving higher accuracy. The XGBoost classifier, because of its speed and higher accuracy, can easily handle higher volumes of data and provide efficient results, thus becoming quite scalable and useful for resolving the issues posed by insider threats. Experimental results clearly indicate XGBoost is better suited for accuracy and yields a result of 98% for complicated scenarios related to threats, while Random Forest and Decision Tree help yield better results and save resources.