Security and Privacy

This paper presents a portable electroencephalography (EEG) monitoring system designed for real-time brain signal acquisition, processing, and secure data transmission. The system integrates eight Ag/AgCl electrodes configured in four channels, using low-noise instrumentation amplifiers (AD620), analog multiplexers (CD4051), and operational amplifiers (TL081) for signal conditioning. An ESP32 microcontroller manages data acquisition, processing, and visualization on a 128\(\,\times \,\)64 I2C OLEDdisplay. Secure data transmission is achieved using the ATECC608A secure element and W5500 Ethernet controller, with a tamper detection switch ensuring physical security. The system is powered by a 3.7V Li-Po battery with an isolated power module(B0505S-1W). The design spans biomedical engineering, electronics, embedded systems, signal processing, IoT, and cybersecurity. Electrode placement follows the 10–20 international system, and the system is compared with state-ofthe-art EEG systems to highlight its unique features, such as portability, security and cost-effectiveness. Detailed subsections cover instrumentation amplifier design, active filters, multiplexer design, electrode placement, hardware implementation, and security measures. Lab tests in a controlled setting showed effective detection of predicted EEG patterns like alpha waves, with minimal artifact and noise. The system maintains a balance among simplicity, performance, and portability and thus is well positioned for use in educational settings, initial neurological investigations, and brain computer interface research. This paper points to the capability of low cost hardware to make EEG technology more democratic and sets the stage for future upgrades, such as wireless connectivity and enhanced signal processing.

Cryptography forms the basis for secure communications in networks that provide mechanisms for the assurance of information confidentiality, integrity, authenticity, and non-repudiation. The steady development of quantum computers—those processing powers strong enough to probably compromise the commonly used forms of digital encryption, basically—poses a whole new potential security threat to network infrastructures and their services widely used today. Theoretically, an ideal quantum computer has the power to break majority of the widely utilized encryption services today, including Rivest–Shamir–Adleman (RSA), Diffie-Hellman (DH), Elliptic Curve Cryptography etc. Thus, the security of these systems is based on computational impracticability as it is observed in the modern world. However, a quantum algorithm such as Shor’s or Grover’s would theoretically speed up the search for symmetric and asymmetric encryption keys, respectively. This would prompt a need for a longer key to maintain the current security level. However, the huge difference should be kept in mind—between quantum computers of today and their problems for public key cryptosystems: they are expected on the horizon. And quantum-resistant algorithms should, meanwhile, be tested and assessed in parallel in various domain areas. This research henceforth explores the integration of quantum-resistant cryptographic algorithms in LTE networks. Sensitive data exchange involved in the LTE handover process and ways to protect them using some kind of quantum-resistant algorithms can be explored.

Wireless Sensor Networks (WSNs) are increasingly vital across domains, yet remain highly susceptible to malicious node attacks which can compromise network reliability and security. This paper presents an interpretable, energy-efficient dual-stage detection framework combining adaptive trust-aware decision trees at the cluster head with server-side hybrid deep and ensemble models. Suspicious cases are effectively elevated for advanced classification using CNN+ RF and AE+ LightGBM, while SHAP-based interpretability provides actionable transparency. A thorough analysis of the SensorNetGuard data indicates that our framework is superior to the previous ML/DL network robotics baselines in terms of performance and stability, imposing a smaller cost on resources, demonstrating an effective learning mechanism suitable for deployment in real-world WSN setups.

Broadcast encryption is a cryptographic method used to send messages securely to multiple users over a broadcast channel. This method is especially useful in Vehicular Ad Hoc Networks (VANETs), where data needs to be transmitted to various vehicles. However, sending different messages to different groups using traditional broadcast encryption can be costly. To address this issue, we use multi-channel broadcast encryption (MCBE) technique. The MCBE scheme we use is dynamic, which makes it easier to add or remove vehicles and Setup phase allows to generate public key and secret key on requirement. We propose a broadcast encryption scheme for VANETs that supports both the dynamic and the inclusive-exclusive property. Existing constructions do not offer the above features. Furthermore, our proposed scheme exhibits constant ciphertext overhead and provides adaptive security.

With the advent of autonomous vehicles, the Vehicular Ad-hoc NETworks (VANETs) are emerging as an important technology for intelligent transportation systems. A privacy-preserving authentication scheme is considered as a promising solution for the evolving attacks in VANET environment. The existing ring-signature based authentication solutions in VANET either lack traceability of the signers or have high computation cost making it unsuitable for practical applications. To address this we propose an authentication scheme which offers both privacy and traceability with computation cost linear in the size of the ring O(n). The other problem is that existing schemes assume the certificate authority to be trusted. But this has serious implication if it is compromised. To address this the proposed work leverages Hyperledger Fabric to achieve transparency of the certificate authority. The network participants can monitor and hence verify the activities of a certificate authority. The performance of proposed work is measured and compared with the existing schemes. A security proof illustrating that the proposed work achieves the security goals of a privacy-preserving authentication scheme is also presented.

Secure group key management (GKM) schemes typically operate under a passive adversary model to evaluate the requirements for forward and backward secrecy. In this model, the adversary can join and leave the group at will. In contrast, the strong adversary model is known as the strong active outsider adversary (SAOA) model, where the adversary is permitted to compromise a legitimate user within the group. Research has shown that all schemes based on the Chinese Remainder Theorem (CRT) are vulnerable to an SAOA adversary. One of the open problems is the design of a secure GKM scheme that can withstand an SAOA adversary without introducing significant complexity. In this work, we propose a first-of-its-kind CRT-based GKM scheme that remains secure against an SAOA adversary while maintaining a rekeying complexity comparable to that of existing CRT-based GKM schemes designed for passive adversaries. Our proposed method can be applied to enhance the security of all current CRT-based key management schemes against SAOA adversaries.

Modern career planning requires relatively adaptive strategies that can keep up with rapid changes in technology, transformations in skill-related requirements, and more and more pronounced non-linear characterizations of career sequences. Fixed traditional recommendation systems cannot grapple with the variety and dynamics of contemporary career paths. This comprehensive review touches upon this very promising hybrid approach wherein multi-agent frameworks operate in conjunction with heterogeneous reasoning paradigms, knowledge graphs, and simulation-based modeling to achieve true personalization and adaptive career strategy development. The analysis synthesizes recent studies cutting across agent decision making, explainable AI (XAI), skill ontology mapping, temporal knowledge graph forecasting, and multi-agent persona-driven systems. Looking across fields, the study highlights noteworthy synergistic benefits, inherent hindrances, and practical opportunities towards linking these into one framework. The hybrid system proposed dynamically produces, evaluates, and perpetually updates career strategies through the interplay of simulation and multi-agent AI technologies. This area has long been the remotest of both fields, holding great promise toward developing more resilient career-planning systems that are truly aware of what the user wants at a certain point in time and that can adapt along with the evolution of the user’s desires and market demands.

A deep learning-based system for detecting counterfeit currency that is specific to Indian banknotes is proposed in this work. We create a custom dataset of real and fake notes across several denominations, use augmentation to address imbalance, and assess transfer learning models (AlexNet, InceptionV3), in contrast to previous works that either use small datasets or lack deployment. InceptionV3 outperformed AlexNet with an F1-score of 97.24% and 99 % validation accuracy. Real-time detection from webcam input or uploaded images is made possible by a Flask-based web application. Our contribution is to provide a lightweight, precise, and easily accessible solution that bridges the gap between research and deployment. The dataset will be expanded, multi-currency support will be integrated, and sophisticated counterfeiting techniques will be addressed in future work.

Next-generation wireless networks, including edge intelligence and wireless distributed learning systems, confront two main obstacles: protecting privacy and communication efficiency. This work addresses these challenges within a distributed learning framework by leveraging the inherent privacy advantages of quantization.

We utilize a Laplace mechanism based on random quantization that simultaneously ensures communication efficiency and robust privacy protection. Unlike existing Gaussian mechanisms, which do not account for decoder or server-level privacy, our approach safeguards against an honest-but-curious server attempting to decode data using dither signals. This mechanism guarantees privacy not only for the database and downstream processes but also at the server level.

Through extensive evaluation in a distributed learning setup, we validate the effectiveness of the Laplace mechanism on datasets such as MNIST and CIFAR-10. Our results demonstrate improved accuracy compared to the Gaussian mechanism while simultaneously ensuring communication efficiency and privacy protection to both the decoder (server) and database through precise realization of the Laplace distribution, which is often a challenge while simultaneously achieving communication efficiency and privacy protection with the help of quantization. This highlights the potential of random quantization in achieving both privacy and utility in distributed systems.

The rapid proliferation of Internet of Things (IoT) devices has introduced unique security challenges, necessitating efficient and lightweight intrusion detection systems tailored to resource-constrained environments. Traditional intrusion detection systems (IDS) often fall short when applied to IoT networks due to their computational complexity and inefficiency in handling high-dimensional data. To address these challenges, this work proposes a lightweight and high-performance intrusion detection framework specifically optimized for IoT networks. The main contribution lies in the integration of the Fisher Score feature selection method combined with a top-k dimensionality constraint, enabling improved model interpretability and classification performance while preserving computational efficiency. Moreover, a unified sampling strategy is employed to address class imbalance in two major IoT botnet variants, Mirai and Gafgyt. The proposed system is thoroughly evaluated using the N-BaIoT dataset, where it achieves an accuracy of 99.92% with the Decision Tree classifier, while other models such as AdaBoost, Gradient Boosting, and Random Forest also demonstrate strong performance. Furthermore, the system’s time efficiency and model simplicity support its feasibility for real-time deployment in practical IoT environments.

The use of machine learning algorithms for anomaly detection has yielded impressive results, leading to significant developments in domains such as healthcare and finance. It is evident that machine learning algorithms require large amounts of data for training, making collaboration essential to obtain sufficient data. However, with such collaboration, the need for privacy whether for individuals or groups has become increasingly important. Numerous instances highlight the negative and potentially dangerous consequences of failing to maintain privacy. Federated learning and privacy-preserving distributed learning techniques ensure privacy when multiple clients or groups collaborate.

This paper introduces a novel privacy-preserving distributed learning framework for anomaly detection, integrating the quorum consensus protocol from distributed systems. This framework is model-agnostic in the sense that it can be applied to any machine learning anomaly detection model that outputs anomaly score. For proof of concept, the Isolation Forest algorithm is employed to detect local anomalies at each client. Empirical evaluations demonstrate that both the equal-weighted and weighted quorum-based global models consistently achieve higher F1 scores compared to individual local models across diverse client configurations. This work presents an effective approach to enhance anomaly detection capabilities in multi-client environments while robustly preserving data privacy.

The increasing technology of digital editing tools has made it harder for humans to distinguish between real and fake faces. Although they can be used in positive applications such as in movies, virtual assistants, video games, and creative arts, some people are using these fake faces for identity fraud, non-consensual content creation, and spreading false information. A primary technology used for generating fake faces is the Generative Adversarial Network (GAN). GANs generate fake faces using adversarial losses between generator and discriminator networks. To minimize this issue, deep learning techniques are being used for distinguishing real and fake faces, achieving more consistent and accurate results. This study introduces a hybrid model that combines the generative strength of Deep Convolutional Generative Adversarial Networks (DCGAN) and the discriminative ability of RESNET50, where DCGAN is one of the advanced GAN technologies to generate new fake faces and RESNET50 is one of the deep convolutional neural networks to classify between the fake faces and real faces. The proposed hybrid model achieves strong performance on the face images dataset, resulting in a precision of 0.91922, a recall of 0.9649, an accuracy of 0.93314, and an ROC under AUC score of 0.986. With the ability to accurately differentiate between real and fake faces, these technologies can help prevent identity fraud, reduce the spread of misinformation, and protect sensitive information from unauthorized access.

Due to digitization, handling big data safely is a big concern nowadays. Cryptography plays a crucial role in protecting sensitive data. Encrypting the data before sending it to the cloud server is a better practice to secure data and perform analysis on encrypted data without revealing raw data. Machine learning is used to make inferences from this encrypted data. The fully homomorphic encryption (FHE) scheme CKKS enables us to perform operations like training, testing, and inference on encrypted data. We have used the TenSEAL library in Python to encrypt our datasets using the CKKS fully homomorphic encryption scheme. As the traditional sigmoid activation function (\(\sigma (x) \)) used in logistic regression is not FHE-friendly, we have designed the logistic regression model to train this encrypted data using various polynomial approximations of the sigmoid. This model is secure, as it has been trained on encrypted data. We have experimented on binary classification datasets and observed diabetes, statlog, titanic, and heart datasets for polynomial approximation (\(0.500781 +0.14670403x +0.001198 x^2 - 0.001006 x^3 \)), which gives better accuracy than the traditional sigmoid activation function.