Praveen Ravula, University of Florida, United States of America
This paper presents the U.S. AI-driven Vulnerability Assessment and Containment Framework (US-AIVAC framework), which can be used to better defend and resilient U.S. critical infrastructure against increasing cybersecurity risks. The suggested system will incorporate intelligent preprocessing, risk assessment through machine learning, automated containment systems to detect and remove vulnerabilities in the infrastructure settings. First, data preprocessing methods (e.g. data cleaning, normalisation, etc.) are done to make the data consistent and reliable to be analysed. To accomplish the vulnerability risk assessment, the Deep Neural Network (DNN) model is used, and vulnerabilities related attributes are identified, and the vulnerability severity are predicted. Moreover, a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network is employed to accommodate spatial and sequential rates of vulnerability data and thus detect intricate cyber-attack behaviours. The suggested US-AIVAC model is expected to enhance the priority of vulnerabilities, enhance threat detection and response to cybersecurity in a timely manner.
AI-Driven Cybersecurity, Vulnerability Risk Assessment, CNN–LSTM Threat Detection, Automated Containment Framework, U.S. Critical Infrastructure Protection.
Sarra Bouzayane, Imène Sekkiou, Houssam Moumouh and Ammar Djebabla Research and Development Department, Caplogy Innovation, Paris, France
Real-time pedestrian trajectory prediction is critical for autonomous drone navigation in indoor environments. While the FORTRAJ system successfully addresses trajectory prediction in fisheye camera environments, its standard configuration requires 8 complete observations before generating predictions, introducing significant activation latency that compromises reactivity during critical early moments. This paper introduces E-FORTRAJ (Early FORTRAJ), an enhanced system that reduces activation latency by 50% through two complementary innovations: (1) enrichment of input trajectories with 12 scientific descriptors spanning kinematic, geometric, statistical, and contextual features to compensate for reduced temporal context, and (2) simple linear extrapolation that generates missing observations while respecting physical constraints. Experiments on the ETH/UCY benchmark demonstrate that E-FORTRAJ maintains competitive performance while triggering predictions from only 4 observations instead of 8. Cross-domain generalization tests on fisheye datasets (HABBOF, Caplogy) confirm the approach's robustness, with E-FORTRAJ achieving ADE of 0.32m and FDE of 0.38m on HABBOF, outperforming both standard and enriched-only configurations. Real-world validation on Caplogy videos at 25 fps shows prediction activation within 160ms (4 frames), demonstrating practical feasibility for drone deployment. These results establish E-FORTRAJ as a viable solution for latency-critical applications requiring early trajectory anticipation in fisheye environments.
Trajectory Prediction, Fisheye Camera, Early Prediction, Feature Engineering, Linear Extrapolation, Autonomous Drones .
Sarra Bouzayane, Yassir Zardoua, Mouad Kahouadji, Imène Sekkiou Research and Development Department, Caplogy Innovation, Paris, France
Fisheye cameras offer wide field-of-view advantages for robotics and autonomous systems, but their inherent radial distortions complicate computer vision tasks. Classical calibration methods require time-consuming manual configuration and must be repeated when intrinsic parameters change. This paper explores deep learning alternatives for direct fisheye-to-perspective image transformation, bypassing traditional calibration. We evaluate two complementary architectures: U-Net with ResNet34 encoder for geometric transformation, and conditional Generative Adversarial Network (cGAN) with Pix2Pix framework for enhanced perceptual realism. Both models are trained on the synthetic Multi-FoV dataset containing 2,500 fisheye-perspective image pairs from urban environments. Experimental results demonstrate successful distortion correction with distinct performance profiles: U-Net achieves Structural Similarity Index Measure (SSIM) of 0.947 emphasizing structural preservation, while cGAN reaches Peak Signal-to-Noise Ratio (PSNR) of 28 dB through adversarial refinement, reducing Mean Absolute Error (MAE) by 82%. We analyze the architectural trade-offs between pixel fidelity and structural coherence, discuss fundamental limitations regarding field-of-view preservation, and propose hybrid multi-projection approaches for future work.
Fisheye Camera Calibration, Deep Learning, U-Net, Generative Adversarial Network, Image Rectification .
Sudhanshu Sekhar Tripathy and Bichitrananda Behera, C.V. Raman Global University, India
Network Intrusion Detection Systems (NIDS) are essential for securing networks by identifying and mitigating unauthorized activities indicative of cyberattacks. As cyber threats grow increasingly sophisticated, NIDS must evolve to detect both emerging threats and deviations from normal behavior. This study explores the application of machine learning (ML) methods to improve the NIDS accuracy through analyzing intricate structures in deep-featured network traffic records. Leveraging the 1999 KDD CUP intrusion dataset as a benchmark, this research evaluates and optimizes several ML algorithms, including Support Vector Machines (SVM), Naïve Bayes variants (MNB, BNB), Random Forest (RF), k-Nearest Neighbors (k-NN), Decision Trees (DT), AdaBoost, XGBoost, Logistic Regression (LR), Ridge Classifier, Passive-Aggressive (PA) Classifier, Rocchio Classifier, Artificial Neural Networks (ANN), and Perceptron (PPN). Initial evaluations without hyper-parameter optimization demonstrated suboptimal performance, highlighting the importance of tuning to enhance classification accuracy. After hyper-parameter optimization using grid and random search techniques, the SVM classifier achieved 99.12% accuracy with a 0.0091 False Alarm Rate (FAR), outperforming its default configuration (98.08% accuracy, 0.0123 FAR) and all other classifiers. This result confirms that SVM accomplishes the highest accuracy among the evaluated classifiers. We validated the effectiveness of all classifiers using a tenfold cross-validation approach, incorporating Recursive Feature Elimination (RFE) for feature selection to enhance the classifiers accuracy and efficiency. Our outcomes indicate that ML classifiers are both adaptable and reliable, contributing to enhanced accuracy in systems for detecting network intrusions.
Machine learning classification systems, Network intrusion detection mechanism, KDD CUP 1999 data repository, Hyper-parameter tuning, Performance evaluation, Classification accuracy .
Nikitha Merilena Jonnada, University of the Cumberlands, USA
Database and system security are critical components in modern information technology, underpinning the reliability and trustworthiness of digital services. With the proliferation of cloud computing, Internet of Things (IoT) devices, and mobile platforms, the attack surface for cyber threats has expanded significantly, creating challenges for confidentiality, integrity, and availability of data. This paper provides a comprehensive review of contemporary database and system security concepts, including access control models, encryption techniques, intrusion detection, and auditing practices. Emerging threats such as ransomware, supply chain attacks, and insider threats are analyzed, alongside mitigation strategies including artificial intelligence (AI)-driven monitoring, blockchain-based integrity verification, and quantum-resistant cryptography. Through case studies in healthcare, finance, and critical infrastructure, the paper highlights practical applications and challenges of security implementation. Finally, it identifies future directions in adaptive security frameworks, zero trust architectures, and privacy-preserving computation, emphasizing the need for a proactive and resilient approach to securing databases and systems.
Database security, system security, access control, encryption, Artificial Intelligence.
Nijat Hasanli, Faculty of Economic Sciences, University of Warsaw, Warsaw, Poland
This paper evaluates whether large language model (LLM)-based sentiment analysis can detect financial distress more accurately than traditional dictionary-based methods. Using the 2023 U.S. bank failures as a natural experiment, Silicon Valley Bank (SVB), Signature Bank, and First Republic Bank each failed during March–May 2023, we construct monthly sentiment indices for five banks using an LLM alongside VADER, TextBlob, and FinBERT under an identical weighting framework. The LLM index consistently declines ahead of and during the failure period for the three distressed institutions while remaining stable for the two control banks (Bank of America, JPMorgan Chase). VADER, TextBlob, and FinBERT fail to detect the distress, remaining strongly positive throughout. Cohen’s Kappa coefficients near zero (0.01–0.22) confirm that the methods capture fundamentally different signals. The LLM index is constructed using a severity-weighted aggregation scheme incorporating source credibility, model confidence, and recency, normalised via a tanh transformation. These findings suggest that LLMs interpret financial context rather than merely counting sentiment-bearing words, offering a meaningful advance for early-warning and financial risk monitoring applications.
LLM, sentiment analysis, financial distress, bank failures, VADER, TextBlob, FinBERT, early-warning systems, Cohen’s Kappa, sentiment index
Kandarp Mukeshkumar Sharda and Aliyu Sani Sambo, University of Wales Trinity Saint David, United Kingdom
Traditional portfolio management systems rely on static rules or fixed prompts, which limit adaptability to dynamic market conditions. This paper proposes a closed loop, multi agent decision system that introduces prompt level learning as a scalable alternative to model retraining. The architecture consists of specialised agents responsible for market signal extraction, sentiment understanding, macroeconomic analysis, risk control, and portfolio construction, coordinated through the DSPy framework and powered by Llama 3.1 8B. A key contribution is a closed-loop optimisation mechanism that continuously refines agent reasoning using real trading outcomes without human intervention: moderate drawdowns trigger incremental prompt updates, while severe drawdowns activate full prompt reconfiguration. Empirical evaluation on a six year dataset (2015–2020) shows the system achieves over 80% cumulative returns with improved risk adjusted performance (Sharpe > 1.5), outperforming a SPY buy and hold benchmark, including during the COVID 19 market disruption. These findings demonstrate the potential of prompt level adaptation for robust, autonomous financial decision systems.
Multi-Agent Systems, Large Language Models, Portfolio Management, Prompt Optimisation, Dynamic Feedback Loop
Vikas Jain, Network Technology Department, India
The exponential growth of 5G deployments and the emergence of 6G research have created unprecedented operational complexity for network operators managing heterogeneous, multi-vendor IoT ecosystems. This paper proposes a vendor-neutral, multi-tenant Network Digital Twin (NDT) platform integrated with AI- driven Operations (AIOps) to address real-time telemetry ingestion, cross-vendor KPI normalisation, predictive anomaly detection, and closed-loop automation at scale. The proposed architecture employs an event-driven microservices model underpinned by a graph-based topology engine, time-series persistence, and ensemble machine learning models. Key use cases including network health twinning, what-if simulation, energy optimisation, and root cause analysis are detailed. The framework is evaluated against operational benchmarks, demonstrating significant reductions in mean-time-to-detect and operational expenditure. The architecture aligns with O-RAN principles and IoT data pipeline requirements, positioning the platform as a foundational layer for autonomous, intelligent network operations.
Network Digital Twin, AIOps, 5G/6G IoT, O-RAN, Predictive Anomaly Detection
Prashant Chaudhary, Pavan Kurariya, Jahnavi Bodhankar and Lenali Singh, Centre for Development of Advanced Computing, India
Machine Translation has become one of the major application areas of Artificial Intelligence (AI) and Natural Language Processing (NLP), especially in multilingual countries like India. Although recent Neural Machine Translation systems have shown good performance for several language pairs, translation quality is still inconsistent for many Indian languages because of linguistic and structural differences between English and Indian language families. Most Indian languages are morphologically rich and contain flexible word order, complex agreement patterns, compound constructions, and context-dependent grammatical forms. Because of this, direct translation from English often produces structurally incorrect or semantically weak output. In many existing systems, the source sentence is passed to the translation model without sufficient linguistic analysis. As a result, ambiguity present in the source text propagates further during translation. This work focuses on the importance of linguistic enrichment before the translation stage. The proposed framework, named Unified Linguistic-Aware Pre-Translation Framework, introduces a coordinated pre-processing layer for English-to-Indian Machine Translation (MT). The objective is not only to clean the input sentence, but also to generate a linguistically stronger intermediate representation that can support downstream translation models more effectively. The framework combines multiple linguistic processing stages including POS refinement, chunk detection, clause boundary analysis, contextual token handling, syntactic structure preparation, and morphology-related processing. Instead of executing these modules independently, the proposed system allows interaction between lexical, syntactic, and morphological information during analysis. This helps reduce structural ambiguity and improves sentence-level interpretation before translation begins. The need for such a framework becomes more relevant in the context of Indian languages where morphology and grammatical relations carry significant semantic information. Small structural errors in the source representation can often affect fluency and meaning preservation in the translated output. Therefore, enriching the source sentence at the pre-parsing level can contribute toward better alignment and semantic transfer. The proposed approach is designed to support both conventional Machine Translation systems and modern AI-based language models. The framework does not replace neural translation architectures; instead, it attempts to strengthen them through explicit linguistic guidance and structured pre-processing. The overall study highlights how classical linguistic analysis can still play an important role in improving multilingual AI systems for Indian languages.
Artificial Intelligent (AI), Natural Language Processing (NLP), Machine Translation (MT)
Leila Habibi, Madjid Maidi, and Boubaker Daachi, LIASD Laboratory, Universit ́e Paris 8 Vincennes–Saint-Denis, France
This paper presents a multimodal vision-language pipeline for object detection and semantic scene understanding based on automatically generated textual prompts. The proposed approach leverages Vision-Language Models (VLMs), including BLIP,BLIP-2, InstructBLIP, and LLaVA, to generate descrip-tive keywords from images, which are then used as queries for open-vocabulary object detectors such as OWLv2 and Grounding DINO. We evaluate the impact of prompt generation on detection performance using a subset of the COCO dataset. Experimental results show that prompt quality significantly affects detection accuracy. InstructBLIP pro-duces consistently strong prompts, while BLIP-2 achieves the best performance when combined with Grounding DINO. While OWLv2 demonstrates more stable and robust performance, Grounding DINO is more sensitive to noisy or complex prompts and tends to generate semantically rich but less standard-ized labels.To address this limitation, we introduce a post-processing strategy that filters, normalizes, and deduplicates predicted labels, leading to substantial improvements in detection performance (F1-score increasing from near zero to above 0.50). The results highlight the importance of integrating prompt engineering and post-processing in multimodal pipelines.This work contributes to the development of interactive vision-language systems for semantic image anal-ysis, with potential applications in educational and human-centered AI systems.
Vision-Language Models, Open-Vocabulary Object Detection, Image Captioning, Prompt Engineering, Multimodal Learning, Grounding DINO, OWLv2
Alaka Krishnan 1 and Rajeev 2 , 1 Master of Computer Application, United Kingdom , 2 Department of Language technology, India
Parts-of-speech (POS) labelling is a fundamental and difficult task for many NLP applications. Malayalam is a Dravidian language belonging to the agglutinative language family of Indian languages. In Malayalam except for other European languages, POS tagging is tagsets that include postposition markers. This process in Malayalam is facing multiple challenges due to its inflectional nature. The Malayalam POS tagset system, developed from the BIS tag set, has several flaws currently being identified. In this study, We discuss a system that contains grammatical categories of Malayalam and the BIS tagset. The main feature of the Malayalam language, which belongs to the highly inflectional language, is the way of adding suffixes after the root. Example: In the Malayalam word ``malajaːɭat̪ːiluɭːoɾakʂaɾam’’, ’malajaːɭam’ (noun), ‘uɭːa’ (conjunction), ‘oɾu’ (adjective), and ‘akʂaɾam’ (noun) with four different meanings are added to the root without dividing them. Mainly in terms of POS tag detection, a computerised system that marks the grammatical categories of Malayalam sentences has not yet been completely developed. The existing POS tagging system is built using machine learning methods and explains how a machine learning model was implemented, addressing the lack of a reliable POS tagging system and the lack of accuracy for Malayalam. Consider this research as a First Benchmark Study and Path Forward for the Under-resourced Language Malayalam. Analysing: How can the resources be improved? The importance of morphological analysis for grammatical categorisation.Here, we used the annotated Malayalam corpus developed by ICFOSS and used this tagged corpus to build the system. We provided detailed analysis with different machine learning algorithms . Our findings demonstrate that TNT tagger and analysing with different algorithms used for the same .
POS (Parts of Speech Tagger), Support vector machine ( SVM), Trigram-n-gram Tagger (TnT), Conditional Random Field (CRF), Large language Model(LLM)
Hasan Abdulrahman , Northern Technical University, Iraq
Accurate drone detection in aerial and ground-based surveillance remains difficult because small unmanned aerial vehicles often appear as low-resolution, low-contrast targets embedded in cluttered backgrounds. Their visual appearance changes rapidly with altitude, viewpoint, illumination, motion blur, compression, and weather, causing many general-purpose detectors to lose fine spatial evidence or produce false alarms from birds, towers, clouds, and building structures. This paper proposes a Multi-Scale Deep Learning Framework (MSDLF) for robust drone detection and classification in complex aerial scenes. The framework integrates a residual convolutional backbone, feature-pyramid aggregation, adaptive scale-attention fusion, decoupled detection heads, and multiscale consistency regularization. Unlike fixed pyramid fusion, the proposed attention module learns image-conditioned weights for features extracted at different strides, allowing the detector to emphasize high-resolution maps for distant drones while preserving deeper semantic context for clutter rejection. Experiments on VisDrone-style aerial imagery and UAV video benchmarks show that MSDLF consistently outperforms representative two stage, one stage, anchor free, and transformer based detectors. The experimental results of comparison, MSDLF achieves 92.1% precision, 90.4% recall, 91.2% F1-score, 94.7% AP50, and 71.9% mAP, improving over YOLOv8 by 3.4 percentage points in precision, 4.2 points in recall, and 4.1 points in mAP. Ablation and robustness studies further show that multi-scale fusion, adaptive attention, and consistency learning each contribute to higher small-object recall and better stability under blur, low light, compression, and scale reduction. These results indicate that scale-aware attention is an effective and practical mechanism for drone detection systems that must balance accuracy, robustness, and near-real-time inference.
Object detection, drone detection, deep learning, multi-scale learning, YOLOv8.
Samsudeen Bankole 1 Yakub Kayode Saheed 2 , 1 Nottingham Trent University, United Kingdom, 2 American University of Nigeria, Adamawa
Retrieval-Augmented Generation (RAG) has emerged as the standard architecture for grounding Large Language Models (LLMs) in factual, domain-specific evidence. In practice, however, the individual components of a modern RAG retriever, dense embeddings, sparse lexical scoring, LLM-driven query expansion, cross-encoder reranking, and adaptive context sizing, do not always compose additively. Under community-default configurations, we show that adding a naïve hybrid step with equal-weighted Reciprocal Rank Fusion degrades NDCG@10 by up to 5.6% on scientific corpora, and that temperature- sampled query expansion with a small LLM degrades it by a further 16.8%. We propose the Adaptive Multi-Stage Vector Retrieval (AMSVR) framework, whose central design principle is a weighted, drift- resistant composition of retrieval signals rather than uniform fusion. AMSVR combines: (i) score- normalised hybrid retrieval with tuneable dense/sparse weights, (ii) deterministic query expansion in which the original query dominates paraphrases, (iii) wide-pool cross-encoder reranking, and (iv) score- variance-driven adaptive Top K selection. On SciFact and NFCorpus we show that a properly configured AMSVR restores monotonic per-stage improvement, delivering +1.4 pp Recall@10 and +1.6 pp NDCG@10 over a strong dense baseline. We further release an offline configuration-diagnostic tool that reproduces our analysis on any new corpus.
Retrieval-Augmented Generation, Dense Retrieval, Hybrid Retrieval, Neural Reranking,Query Expansion, Adaptive Top K, BEIR Benchmark, Large Language Models
Dhuha AL-ADHAMI 1 , Hamza GHARSELLAOUI 2, Olfa Belkahla DRISS 3 ,1 Faculty of Sciences of Tunis, University of Tunis El Manar, Tunis, Tunisia 2 National School of Computer Science (ENSI), University of Manouba, Manouba, Tunisia , 3 ESCT, University of Manouba, Manouba 2010, Tunisia
We propose a content-adaptive image steganography framework using the Coati Optimization Algorithm (COA) with a transformer-based self-attention cost model for statistically secure covert communication. Existing steganography methods embed secret data independently of the local image content, and are vulnerable to modern steganalysis detectors. We improve on this using a non-local attention cost model (parameterized by texture-aware query, key and value projections) trained with COA to minimize the SRM steganalysis footprint over a training set of cover images. Bit allocation is then performed according to the optimized cost map to drive secret bits to perceptually safe locations, and Hamming(7,4) error-correcting codes are used to guarantee reliable recovery of the payload in the presence of channel attacks. We compare the method to LSB, DCT and PVD baselines across the metrics of imperceptibility (PSNR, SSIM), robustness (BER with JPEG, Gaussian, salt-and-pepper, rotation and median-filter attacks), embedding capacity and steganalysis detectability. An ablation study justifies the importance of both the COA training stage as well as the transformer attention mechanism. We demonstrate convergence of COA against PSO, GA and WOA, showing superior fitness optimization performance. Statistical significance is shown through Wilcoxon signed- rank tests.
image steganography; covert communication; Coati Optimization Algorithm; transformer self-attention; content- adaptive embedding; steganalysis security; PSNR; SSIM; error correction; meta-heuristic optimization
Dhuha AL-ADHAMI 1 , Hamza GHARSELLAOUI 2, Olfa Belkahla DRISS 3 ,1 Faculty of Sciences of Tunis, University of Tunis El Manar, Tunis, Tunisia 2 National School of Computer Science (ENSI), University of Manouba, Manouba, Tunisia , 3 ESCT, University of Manouba, Manouba 2010, Tunisia
Image steganography is a critical technique for covert communication, aiming to embed secret data within digital images while preserving visual imperceptibility and minimizing statistical detectability. The Least Significant Bit (LSB) substitution method, despite its simplicity and high embedding capacity, suffers from suboptimal pixel selection that introduces perceptible distortion and vulnerability to steganalysis. This paper proposes a novel Hybrid Particle Swarm Optimization and Meerkat Clan Optimization (PSO-MCO) algorithm to determine the optimal set of pixels for LSB embedding, minimizing distortion while maintaining high embedding capacity. The hybrid mechanism combines PSO's global exploration strength with MCO's local refinement capability through an adaptive switching strategy governed by a convergence velocity threshold. Experiments conducted on five standard 512x512 benchmark images demonstrate that the proposed method achieves a mean Peak Signal-to-Noise Ratio (PSNR) of 51.2 dB, a Bit Error Rate (BER) of 0.0023, and a Structural Similarity Index (SSIM) of 0.9981, outperforming standard LSB, PSO- only, and Meerkat-only approaches by 21.6%, 14.3%, and 13.0% in PSNR, respectively. The results confirm that the hybrid optimization significantly enhances imperceptibility without sacrificing embedding capacity.
Image Steganography, LSB Substitution, Particle Swarm Optimization, Meerkat Clan Optimization, PSNR
