Volume 14 | Issue 7 |

 Abstract

This paper presents an AI-based multimodal system for truth and lie detection that combines audio and video analysis within a unified Flask based framework for non-intrusive deception assessment. Traditional lie detection techniques, such as polygraph tests, are often intrusive, expensive, and not always reliable in real-world scenarios. To overcome these limitations, the proposed system makes use of recent advancements in computer vision and speech processing to analyze both facial behavior and vocal characteristics. The audio module extracts deep speech features along with handcrafted features such as pitch, energy, and zero-crossing rate, while the video module uses a CNN-LSTM architecture to capture spatial and temporal patterns from facial expressions and movements. The outputs from both modules are combined using a multimodal fusion approach to produce a final prediction along with a confidence score. The system supports both file-based inputs and real-time recording through a user-friendly web interface, and it also maintains a history of predictions for better usability. Although previous studies show that multimodal approaches generally perform better than single-modality systems, challenges such as limited datasets, generalization issues, and ethical concerns still remain. Overall, the proposed system provides a practical and scalable approach for developing AI-assisted deception detection applications.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Audio classification, CNN-LSTM, Computer vision, Deception detection, Multimodal learning

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Explainable AI, Chest X-Ray, Grad-CAM, SHAP, RadBERT, DenseNet-121, Multimodal, Clinical NLP.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

This paper proposes a novel approach for real-time fitness evaluation by employing computer vision and deep learning technologies. Existing fitness monitoring systems typically rely on human supervision or wearable sensors, which are often invasive, expensive, and inaccessible for many users. The proposed system uses pose estimation to extract skeletal keypoints from video input and processes them as temporal sequences using Bidirectional Long Short-Term Memory (BiLSTM) models. An automatic labeling strategy based on joint angle thresholds is employed to classify exercise stages, and a multi-model strategy is used to independently handle push-ups, squats, and pull-ups. The system combines classification and detection approaches to accurately identify exercise types and count repetitions by detecting transitions between motion stages. Experimental results demonstrate high accuracy in exercise classification, stage detection, and repetition counting, confirming the potential of the proposed approach as a non-intrusive and efficient solution for real-time fitness evaluation.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

pose estimation, BiLSTM, fitness monitoring, exercise recognition, computer vision, deep learning

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Modern enterprise systems rely extensively on API keys, access tokens, database credentials, and configuration secrets. Such credentials often become compromised via various means like misconfiguration, insider misuse, open-source code exposure, phishing, and malware, with credential-based intrusion becoming one of the most rapidly growing threat vectors in contemporary cybersecurity. Existing honeypot solutions for identifying such threats by deploying static fake credentials have proven to be less and less effective, as they can easily be detected and bypassed by more advanced and experienced threat actors. In this paper, we introduce a Generative AI Honeytoken Factory-a state-of-the-art artificial intelligence-based cyber deception solution, capable of generating and using fake credentials in the form of honeytokens. The proposed solution is based on four computational modules: Generative Honeytoken Factory built upon the combination of the hybrid Diffusion Model and Variational Autoencoder (VAE) with adversarial training, which can generate honeytokens with 90-95% structural accuracy; Attack Behavior Graph Intelligence using Temporal Graph Neural Networks (TGNN) with 90-98% accuracy in detecting attacks; Adaptive Honeytoken Deployment Engine using Multi-Agent Deep Reinforcement Learning (SAC) with 85-96% effectiveness; and Threat Attribution Engine using Heterogeneous Graph Transformer (HGT) with 85-92% attribution accuracy.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Honeytoken; Generative AI; Cyber Deception; Variational Autoencoder; Diffusion Model; Temporal Graph Neural Network; Deep Reinforcement Learning; Threat Attribution; Attack Detection; Credential Security

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Modern enterprise software systems rely heavily on database-driven applications that require continuous manual effort to develop, maintain, and update CRUD (Create, Read, Update, Delete) interfaces and backend APIs whenever the underlying schema evolves. This repeti- tive development cycle is costly, error-prone, and inaccessible to non-technical stakeholders. This paper presents IntelliCRUD, an AI-assisted dynamic data management platform designed to eliminate manual interface development by automatically analysing any MySQL database schema and generating corresponding React-based user interfaces and RESTful backend APIs at runtime.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Dynamic CRUD, Schema-driven UI, Role-Based Access Control, Text-to-SQL, Large Language Models, Machine Learning, Enterprise Data Management, Natural Language Querying, Form Type Inference.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Automatic musical audio to symbolic representation is a challenging task, especially for Carnatic music because of microtonal variations and ornamentations. This work sys- tematically generates Carnatic swaras from polyphonic audio recordings that are mostly based on a major scale structure. The approach combines deep learning-based source separation using Demucs [2] and pitch estimation using CREPE [1], along with octave-invariant pitch-class representation and histogram- based tonic detection [7]. A gamaka-aware smoothing strategy is introduced to enhance stability and reduce frame-level noise. The system outputs swara sequences suitable for music analysis and educational applications. Experimental observations indicate that the proposed approach provides reliable tonic detection and accurate mapping of swaras under controlled conditions

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Pitch Estimation, Tonic Detection, Swara Esti- mation, Tonic Estimation, MIR

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Conventional image compression applies nearly uniform quality across all regions of an image, in contrast to human perception and downstream vision tasks which are much more sensitive to distortions in semantically important regions than in smooth or visually unimportant background areas. The proposed paper addresses this inefficiency through a saliency guided image compression framework that estimates the importance of pixels using three complementary modules: deep salient object detection, semantic object segmentation, and spectral residual saliency analysis [1]. The output of these modules are fused into a unified importance representation, which is then transformed into a spatially varying bit allocation map through an Ascending Cosine Roll down (ACRD) transfer function that emphasizes perceptually relevant regions while suppressing background detail.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

saliency detection, image compression, bit allocation, context aware coding, layered reconstruction.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Aphasia diagnostic assessments are often time-intensive and subjective. While automated tools exist, traditional cascaded architectures, which often rely on Automatic Speech Recognition (ASR) followed by Natural Language Processing (NLP), frequently suffer from cumulative propagation errors, discarding vital prosodic features necessary for precise clinical diagnosis. This thesis introduces NeuroPheno, an innovative end-to-end multimodal diagnostic framework that identifies Aphasia subtypes directly from raw speech signals. By leveraging the Qwen2-Audio architecture, our model processes audio and text inputs simultaneously, eliminating the need for intermediate transcription and preserving critical diagnostic markers. Experimental results demonstrate that this end-to-end approach significantly outperforms traditional cascaded systems, achieving an accuracy of 89.65% and an F1-score of 0.92 on the APROCSA dataset. By bypassing the transcription layer, the model captures subtle semantic and acoustic anomalies frequently overlooked in standard clinical pipelines. These findings suggest that our framework offers a precise, objective, and scalable solution for clinical telemedicine, enabling early diagnosis and continuous, reliable monitoring of communication disorders.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Aphasia, End-to-End Deep Learning, Qwen2-Audio, Multimodal Systems, Clinical Diagnostics, Speech Processing.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Multi-agent reasoning has emerged as a promising approach to improve the reliability of large language models by enabling collaborative problem solving. However, existing methods often rely on loosely co- ordinated interactions, leading to shallow reasoning, unverified assumptions, and propagation of errors across agents.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Learning to Scrutinize: Adversarial Multi-Agent Reasoning with Generator-Scrutinizer Architecture

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The financial services sector demands accurate, scalable, and unbiased mechanisms for loan eligibility assessment and insurance claim verification. Both processes are data-intensive and decision-critical, making them ideal targets for intelligent automation using machine learning and natural language processing. Existing automated loan prediction systems, including the Preferential Selective-Aware Graph Neural Network (PSAGNN) and the Interbank Credit Prediction (ICCP) model, addressed credit rating forecasting but suffered from significant limitations. Existing systems suffer from narrow prediction scope, slow accuracy gains, poor scalability, reliance on traditional techniques, and lack identity verification and insurance claim analysis capabilities. This paper proposes LoanOracle, a Machine Learning- Based Loan Approval and Management System integrated with a Retrieval-Augmented Generation (RAG) Chatbot for Insurance Claim Verification. The loan prediction engine employs a Voting Classifier ensemble combining Logistic Regression, Decision Tree, Random Forest, and Extremely Randomized Trees using soft voting, which averages predicted class probabilities across all estimators to produce the final eligibility decision. The RAG pipeline processes uploaded documents through LangChain loaders, segments them into 500-character

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

machine learning, loan eligibility, retrieval-augmented generation, insurance claim verification, FAISS, LangChain.

  License

Creative Commons Attribution 4.0 and The Open Definition