Rajagiri School of Engineering & Technology

Video steganography has become an essential technique for secure communication, watermarking, and copyright protection in the face of increasing cyber threats and surveillance. Unlike encryption, which often draws attention, steganography embeds sensitive information within multimedia content. Videos, with their large size and redundancy, offer unique opportunities for covert data transmission. While several surveys exist, most either focus broadly on multimodal steganography or provide only limited coverage of video-based methods. They often lack a systematic comparison across spatial, transform, and compressed domains, insufficiently analyse trade-offs among imperceptibility, robustness, and embedding capacity, and rarely address practical constraints such as computational cost, dataset limitations, and resistance to modern steganalysis. This review fills these gaps by offering a taxonomy driven analysis of video steganography techniques. Methods such as Least Significant Bit (LSB), Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), motion vector modification, and hybrid approaches are critically compared using standardized performance metrics. Beyond cataloguing techniques, we evaluate vulnerabilities, highlight limitations in existing research, and propose a roadmap for future developments. Future perspectives emphasize deep learning based adaptive embedding, blockchain-enabled authentication, and optimization strategies for enhanced efficiency and scalability. Thus, this paper contributes a critical, domain specific roadmap that advances secure video steganography beyond the scope of earlier surveys.

Brain Computer Interfaces (BCIs) that interpret neural signals into natural text have the potential to improve communication ability for individuals with speech disabilities. Despite early successes in this field, creating intelligible text from non-invasive EEG signals remains challenging due to the low SNR, high dimensionality of the EEG dataset across time, and the inherent non-stationarity in neural data. Current methods do not fully preserve the full meaning of speech because they rely on simple techniques like linear regression or phonetic mapping, which tend to miss the subtle meanings and context found in natural language. This paper presents EEG to Text, a contrastive latent space guided transformer architecture framework to reduce the semantic gap and increase the ability to reinterpret the relationship between neural patterns and text. EEG to Text creates a shared space that connects EEG signals and text data. It uses a strong, pre-trained text encoder as a guide to help align EEG features over time with the meaning of natural language. This creates a model that can identify meaningfully interpretable neural activity corresponding to linguistic meaning.

Industrial environments, particularly oil refineries, depend on a huge amount of technical documentation, such as operating manuals, safety procedures, and standard operating guidelines, to ensure safe and efficient day-to-day operations. It is essential to find the right information from these documents for smooth operations and to help employees perform their tasks effectively. Manual search methods often require considerable time and can lead to delays, errors, and reduced efficiency. This paper presents a domain-specific offline intelligent assistant built using Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) for refinery knowledge management. The system is designed to operate entirely offline, ensuring strict data confidentiality while delivering accurate and context-sensitive responses. The architecture integrates semantic document processing, hybrid retrieval mechanisms, and optimized LLM inference. The experimental results demonstrate improved retrieval accuracy and response relevance compared to the standalone LLM approaches. The proposed system highlights the feasibility of deploying intelligent systems based on the foundation-model in secure industrial environments.

HVAC systems are a major contributor to energy consumption and carbon emissions in modern buildings, creating the need for intelligent and scalable control methodologies. Even though the performance of Model Predictive Control (MPC) is better, it relies heavily on accurate system models and necessitates computationally intensive online optimisation to limit its applicability in practice. Alternatively, model-free reinforcement learning exists, but the need to explore online makes the approach unsafe in the actual HVAC facilities. This paper presents a carbon-conscious offline reinforcement learning model based on Conservative Q-Learning (CQL) to learn the best HVAC control policy with a historical operation record that lacks interaction in real time. The framework uses behaviour cloning to initialise stable policy, normalised state representations, and reward scaling to improve convergence and robustness. These value-function regularisation methods employ conservative regularisation to discourage overestimation bias in distributionally constrained policy learning. Empirical testing on a multi-objective HVAC data set indicates up to 9.99% improvement in cumulative reward compared with the operation of a baseline HVAC, whilst maintaining a predictable and comprehensible control behaviour. This means that the suggested framework enables scalable, safe, and carbon-efficient HVAC control, which can be deployed in real life.

The rapid increase in digital educational resources has created a need for scalable, intelligent solutions to transform unstructured learning material into structured, interactive assessments. This paper presents EDU-DELL, an AI-powered educational platform designed for automated content extraction, summarization, and assessment generation. The system supports multi-format inputs, including PDF, DOCX, PPTX, TXT, and web URLs. It integrates transformer-based models such as BERT, FLAN-T5, and Google’s Gemini with spaCy pipelines and semantic search using ChromaDB. EDU-DELL automatically generates multiple-choice questions (MCQs), fill-in-the-blank (FIB) exercises, true/false statements, and concise summaries, ensuring high contextual accuracy. Evaluation on diverse real-world educational datasets demonstrates notable performance in question quality, distractor diversity, and summarization coherence. The platform offers practical value for educators, learners, and institutions by reducing manual workload, enabling self-assessment, and improving content accessibility. These results position EDU-DELL as a robust, scalable solution for modern AI-driven educational technology.