International Journal of Information and Computer Security (9 papers in press)

Regular Issues

• Code vulnerability detection based on improved GNN and attention mechanism  
  by Cheng Yang, Mingsheng Yang 
  Abstract: This study proposes an improved graph convolutional network (GCN) model that integrates a dynamic inter-layer attention mechanism. The core of this model is as follows: firstly, a dynamic inter-layer attention mechanism is designed. Through a learnable weight allocation unit, it can adaptively integrate the graph features at different depths of the model. Secondly, a lightweight method based on node importance scoring and dynamic parameter pruning is introduced to improve the efficiency of the model. Finally, a graph transformation encoder is combined to enhance the interpretability of the models decision-making process. The comprehensive detection accuracy of the proposed model reaches 97%, and its performance is significantly superior to the mainstream benchmark models. The model was evaluated on a mixed dataset containing a standard vulnerability dataset (including a list of five common defects) and real project code. This provides an efficient and reliable solution for automated code security auditing.
  Keywords: graph neural network; GNN; self-attention mechanism; SAM; code vulnerabilities; vulnerability detection; interpretability; graph structure.
  DOI: 10.1504/IJICS.2026.10079451
   
  
• Generation of believable fake combinational logic circuits for cyber deception  
  by Nilin Prabhaker, Ghanshyam S. Bopche, Michael Arock 
  Abstract: The exponential increase in intellectual property (IP) theft has raised concern for organisations and government entities. Existing research suggests the use of fake document for protection against IP theft. To protect the logic circuit present in the IP document, we propose believable fake combinational logic circuits generation engine (BFCLCGE), which is an extension of our previous work (FBLCGE). We addressed the exhaustive search problem by modifying the original Boolean equation and introducing replacement constraints, thereby reducing overall time complexity. Additionally, we overcame the limitation of FBLCGE by incorporating ESPRESSO algorithm for efficient minimisation of Boolean equations with a large number of input variables. To test the believability of generated fake Boolean circuits, we implemented a virtual display (16-segment) and simulated the outputs. The proposed approach is efficient and generates indistinguishable fake circuits, which can be used to create believable fake documents for cyber deception against potential IP theft.
  Keywords: intellectual property; data ex-filtration; combinational logic circuits; hardware security; cyber security; cyber deception; decoy logic circuits; decoy documents.
  DOI: 10.1504/IJICS.2026.10079452
   
  
• Differentially private influence maximisation in social networks  
  by Sri Satya Monica Bandaru, Nagesh Bhattu Sristy, Karthick Seshdri 
  Abstract: Influence maximisation (IM) aims to identify influential individuals or nodes in a social network for information transmission. IM finds key applications in targeted advertising, public health initiatives, and viral marketing. Widely used diffusion-based methods for IM lack generalisation. Recent research incorporates node embeddings for improving seed user identification. However, network embedding methodologies often reveal sensitive information about individuals. This work proposes a task-aware influence maximisation approach that integrates formal differential privacy guarantees with adversarial graph representation learning. By jointly learning source and destination node embeddings, we create a direction-aware generative adversarial network (GAN) that effectively captures asymmetric influence interactions in directed social networks. Noise-perturbed optimisation is used to guarantee differential privacy during the embedding learning step, protecting sensitive information regarding relationships. A CELF-based greedy IM algorithm allows influence spread to be optimised directly from differentially private representations, allowing for the efficient identification of influential nodes while maintaining privacy. Experimental evaluation on real-world networks demonstrates a 30% increase in privacy protection against link inference attacks, along with comparisons to traditional IM techniques.
  Keywords: social network analysis; SNA; influence maximisation; differential privacy; diffusion probabilities; node embeddings; moment accountant.
  DOI: 10.1504/IJICS.2026.10079453
   
  
• Secured sharing of patients electronic health records with modified proxy blind signature scheme enabled hybrid authentication  
  by Ankita Vasant Pande, Sheetal S. Dhande 
  Abstract: Consent management in electronic health record systems (EHRS) plays a crucial role in ensuring the privacy and security of patient data sharing. Existing research has primarily focused on various consent models but often faces drawbacks such as complexity in implementation, lack of robustness in authentication mechanisms, and issues with scalability and efficiency. The research proposes a novel approach named Modified Proxy Blind Signature Verification with Hybrid Authentication for Consent Management (mPBS-HA CM) in EHRS that overcomes the existing limitations and provides an efficient outcome. The proposed model integrates hybrid authentication mechanisms to achieve hybrid authentication by enhancing both security and efficiency. The involved hybrid authentication mechanisms ensure secure key exchange and digital signature generation along with providing robust encryption of sensitive data. The research model provides the advantages of enhanced security through mPBS that improved scalability due to efficient techniques, and simplified implementation compared to traditional approaches. The research model with improved performance is evaluated with encryption time, decryption time, transaction time, genuine user rate (GUR), gas usage, memory usage, and responsiveness that achieved 3.475 s, 3.453 s, 1.017 s, 174.872 KB, 326.072 KB, 0.72, and 10.122 ms respectively.
  Keywords: consent management; electronic health records; hybrid encryption; hybrid verification; proxy blind signature scheme; health record security.
  DOI: 10.1504/IJICS.2026.10079470
   
  
• Automatic fake news detection using an optimised transformer with deer hunting optimisation algorithm  
  by Yamini Devi Jonnala, J. Sirisha Devi 
  Abstract: Fake news has become a significant issue, influencing public opinion and decision-making across various platforms. Traditional machine learning algorithms often struggle to accurately detect subtle contextual details within misleading information. Transformer-based models, with their self-attention mechanisms, are particularly well-suited for this task, as they capture deeper contextual relationships within the text. This study proposes an advanced fake news detection system utilising a transformer encoder model, focusing on textual analysis for precise classification. By leveraging the transformers ability to identify complex patterns and contextual nuances, the system improves the detection of misleading or false information. To further enhance model performance, hyperparameter optimisation is applied using the deer hunting optimisation algorithm (DHOA). The system is evaluated on a benchmark dataset, demonstrating its effectiveness with high recall and precision metrics. The proposed approach addresses the escalating issue of misinformation by combining cutting-edge natural language processing with an innovative optimisation technique. For the proposed method, transformer encoders were used, achieving an accuracy of 98.71%, a recall of 98.7%, and an F-measure of 98.8% on dataset 1. on dataset 2, the model achieved an accuracy of 98.23%, a recall of 98.24%, and an F-measure of 98.36%.
  Keywords: transformer encoders; deer hunting optimisation algorithm; DHOA; fake news; hyperparameter.
  DOI: 10.1504/IJICS.2026.10079776
   
  
• W-state and chaotic mapping-driven: a new three-party key agreement protocol  
  by Ningning Xu, Haocheng Kan, Peng Guo, Hongfeng Zhu 
  Abstract: In three-party communication scenarios, the high cost of quantum devices poses a significant barrier to the widespread adoption of quantum key agreement technologies. To address this challenge, we propose a novel semi-quantum key agreement (SQKA) protocol. This protocol integrates W-states and chaotic sequences to simultaneously prevent information leakage and malicious tampering during communication. The core innovations of the proposed scheme are threefold: first, the pseudo-randomness of chaotic sequences is introduced into quantum communication, which ensures the security of classical channel transmission while generating non-unique sequences for enhanced flexibility; second, the strong entanglement property of W-states is leveraged to enable timely detection of channel attacks; third, a two-way verification mechanism between communicating parties is incorporated to further strengthen interaction security. The security assessment indicates that this protocol can resist both internal and external attacks. Finally, compared with the communication efficiency of other schemes, our scheme can still maintain a very high efficiency of 8.7%.
  Keywords: semi-quantum; W-state; chaotic sequence; key agreement.
  DOI: 10.1504/IJICS.2026.10079777
   
  
• DEARNN: a hybrid deep learning approach for cyberbullying detection in Twitter social media platform  
  by Nagaraju Sonti, T.V.L. Pranathi, S. Alekhya, T. Haritha, M.L. Illisha 
  Abstract: Cyberbullying is becoming a more widespread problem on social media, especially on sites like Twitter where identification is made more difficult by colloquial language and concealed connotations. In order to detect dangerous texts, this study presents dual encoder attention-based recurrent neural network (DEARNN), a sophisticated deep learning model. DEARNN uses an attention mechanism to highlight important passages in the text that are suggestive of bullying and uses recurrent neural networks (RNNs) to assess word sequences. Furthermore, word embeddings improve the models capacity to recognise intricate patterns. By identifying both overt and covert kinds of bullying, DEARNN outperforms conventional techniques, according to testing on huge Twitter datasets. In order to increase accuracy, future improvements will concentrate on adding multimedia material and broadening language support.
  Keywords: cyberbullying detection; recurrent neural networks; RNNs; deep learning; language model.
  DOI: 10.1504/IJICS.2026.10079806
   
  
• Context-aware differential privacy for federated learning in healthcare  
  by Kalyani Pampattiwar, Namrata Patel, Masooda Modak 
  Abstract: Rising interest in collaborative medical artificial intelligence has encouraged adoption of federated learning to protect patient data. Yet many models add uniform differential privacy noise, ignoring patient age, disease severity and institutional risk, which weakens the trade-off between privacy and model utility in heterogeneous healthcare data. We present a context-aware differential privacy framework for federated learning that provides patient-level, context-specific privacy through five modules: patient-aware noise calibration network assigns noise from demographics and disease factors disease-specific gradient clipping sets complexity-adjusted thresholds federated context embedding exchange shares non-identifiable context embeddings adaptive contextual utility estimator predicts local utility loss and adjusts noise hierarchical privacy budget scheduler allocates budgets across patients, hospitals and disease cohorts. On multi-institutional healthcare datasets, the framework improves privacy efficiency by 25%, retains over 95% baseline accuracy and increases robustness under non-independent and identically distributed conditions, enabling practical, privacy-adaptive federated systems for clinical decision support at scale.
  Keywords: federated learning; differential privacy; healthcare AI; patient-specific privacy; context-aware privacy; process.
  DOI: 10.1504/IJICS.2026.10079981
   
  
• Ensemble learning for anomaly detection: a comparative study of fusion of classical and neural models in cybersecurity logs  
  by Mrityunjay Brahma, Hemanta Kumar Kalita 
  Abstract: In the face of increasingly sophisticated cyber threats, traditional signature-based detection systems often fall short, especially against zero-day attacks. This research presents a hybrid anomaly detection framework that integrates classical supervised machine learning algorithms and deep learning models to enhance detection accuracy while minimising false positive rates. Using the UNSW NB15 proxy server dataset, we evaluate various classical machine learning models k-nearest neighbours, decision trees, Naive Bayes, logistic regression, random forest, and gradient boosting alongside neural networks like multi-layer perceptron and long short-term memory. Two ensemble based models (M1 with multilayer perceptron, M2 with long short-term memory) are designed using soft voting classifiers, combining algorithm subsets in odd-numbered groupings (3, 5, 7) to ensure clear majority decisions. Our findings demonstrate that the ensemble approach significantly improves classification performance, with model M1s largest ensemble (seven algorithms) achieving up to 99% accuracy and with lowest 1% false positive rate. This study underscores the effectiveness of combining machine learning and neural networks/deep learning methods for adaptive, scalable, and precise network anomaly detection systems, providing a robust defense against evolving cybersecurity threats.
  Keywords: classical supervised learning; cyber security; deep learing; ensemble model; neural network.
  DOI: 10.1504/IJICS.2026.10079982