Publications

HERALD protects selected sensitive tokens with client-side deterministic ciphertext while preserving context and downstream clinical utility.

While large language models (LLMs) are increasingly used for clinical applications, many existing pipelines require sending raw sensitive health information to remote servers, heightening the risk of privacy leakage. Encrypting an entire dataset introduces prohibitive computational, alignment, and communication overheads. To preserve privacy while maintaining usability, we present Healthcare Encryption & Redaction via Adaptive Linguistic Decomposition (HERALD), a token-level cryptographic redaction framework designed to encrypt only sensitive tokens while preserving surrounding context for downstream model utility. HERALD combines medical named-entity recognition with part-of-speech-driven policies to select candidate tokens, performs targeted lemmatization to stabilize surface forms, and substitutes each protected token with deterministic ciphertext wrapped in explicit delimiters. The model-agnostic framework operates entirely on the client side, keeping sensitive content encrypted during storage, transmission, and processing without changes to downstream models. Evaluation on public classification and medical question-answering datasets shows that fully secured baselines suffer substantial utility loss, whereas HERALD consistently recovers performance close to plaintext. This record remains under review and is not presented as accepted or published.

Eight classifiers are compared across Mendeley, tokenized-URL, and merged data, with the merged dataset producing the strongest overall detection results.

This paper presents a discrete mathematical modelling of cybersecurity phishing attack detection methodologies, emphasizing the need for continual advances in detection. Leveraging mathematical modelling and machine-learning algorithms, the study evaluates Logistic Regression, k-Nearest Neighbors, Support Vector Machines, Random Forest, Gradient Boosting Machines, Neural Networks, CatBoost, and XGBoost on Mendeley, URL-tokenized, and merged datasets. In the original Mendeley dataset, XGBoost achieves 97.24% accuracy, while CatBoost leads after preprocessing with 97.28% accuracy. Tokenized URLs reach a highest accuracy of 91.95%. The combined dataset produces the highest overall accuracy of 97.68%, with SVM excelling in sensitivity and specificity and XGBoost in precision.

BERT-family language models classify phishing URLs strongly on rich features, while an external feature-scarce test exposes a meaningful recall gap.

This study explores the efficacy of four Large Language Models (LLMs)—BERT, DistilBERT, RoBERTa, and DeBERTa—in classifying URLs as either legitimate or phishing. The methodology covers dataset processing, model fine-tuning, and performance evaluation on a primary dataset with extensive features and an external dataset with minimal features. The models consistently achieve high performance on the primary dataset, with AUC scores of 0.99. DistilBERT records F1-score 99.992%, accuracy 99.991%, and precision 99.985% on the primary evaluation. On an external test dataset containing 450,176 labelled URLs, BERT and DeBERTa show low F1-scores of 1.14% and 2.78%, while DistilBERT reaches 42.44% and RoBERTa 68.03%, exposing the sensitivity of results to feature availability.

RecruitView contributes a 2,011-clip multimodal interview dataset and CRMF, a geometry-aware model that improves correlation while using fewer parameters.

Automated personality and soft skill assessment from multimodal behavioral data remains challenging due to limited datasets and methods that fail to capture geometric structure inherent in human traits. We introduce RecruitView, a dataset of 2,011 naturalistic video interview clips from 300+ participants with 27,000 pairwise comparative judgments across 12 dimensions: Big Five personality traits, overall personality score, and six interview performance metrics. To leverage this data, we propose Cross-Modal Regression with Manifold Fusion (CRMF), a geometric deep learning framework that explicitly models behavioral representations across hyperbolic, spherical, and Euclidean manifolds. CRMF employs geometry-specific expert networks to capture hierarchical trait structures, directional behavioral patterns, and continuous performance variations simultaneously. An adaptive routing mechanism dynamically weights expert contributions based on input characteristics. Through principled tangent space fusion, CRMF achieves superior performance while training 40–50% fewer trainable parameters than large multimodal models. Extensive experiments demonstrate that CRMF substantially outperforms the selected baselines, achieving up to 11.4% improvement in Spearman correlation and 6.0% in concordance index. The RecruitView dataset is publicly available at https://huggingface.co/datasets/AI4A-lab/RecruitView. This record remains labelled under review; the arXiv source does not establish acceptance.

ORBIT combines cross-attentive multimodal fusion, language adversaries, and spherical-hyperbolic learning for zero-shot cross-lingual Alzheimer detection.

In this work, we study zero-shot cross-lingual speech-based Alzheimer’s disease detection (SADD). We hypothesize that learning language-invariant multimodal representations by fusing multilingual speech and text pretrained models is essential for reliable transfer to unseen languages, while adversarial learning suppresses language-specific confounds. Empirical results in zero-shot cross-lingual evaluation show that multimodal fusion consistently outperforms unimodal baselines. We propose ORBIT, a framework combining cross-attentive fusion, multi-tap language adversaries, and complementary spherical-hyperbolic geometric learning with consensus clustering. Across settings, ORBIT achieves the strongest performance compared with unimodal models and simple concatenation-based fusion baselines. This work is recorded as accepted to Interspeech 2026; its arXiv record remains the accessible source.

Uc-PrUn couples zero-shot uncertainty estimation with selective unlearning to improve calibration and downstream clinical VLM performance.

In this study, we introduce Uc-PrUn, a principled framework designed to improve the reliability of Vision–Language Models (VLMs) in clinical decision-support systems. The first stage focuses on Bayesian-inspired zero-shot uncertainty quantification using Monte Carlo dropout, while the second stage introduces an uncertainty-aware machine-unlearning strategy. Leveraging the Harvard-FairVLMed dataset, which comprises paired SLO fundus images and clinical notes for glaucoma detection, we evaluate VLMs to quantify epistemic uncertainty and identify high-variance training samples. The pruning and unlearning mechanism selectively removes uncertain samples to enhance model calibration and downstream performance. Experiments show that Uc-PrUn reduces predictive uncertainty and yields consistent gains in accuracy and F1 scores across multiple VLMs, supporting uncertainty-aware pruning in medical AI pipelines.

CNN transfer-learning models classify normal, COVID-19, and pneumonia chest X-rays while AES and fully homomorphic encryption protect the data.

This research presents a privacy-preserving approach to chest X-ray image classification using deep-learning techniques integrated with Advanced Encryption Standard (AES) encryption and Fully Homomorphic Encryption (FHE). The methodology classifies normal, COVID-19 infected, and pneumonia infected images without decrypting the data and evaluates ResNet50, ResNet101, MobileNetV3 Small, MobileNetV3 Large, DenseNet121, and DenseNet161 after preprocessing, hyperparameter selection, and transfer-learning fine-tuning. DenseNet161 achieves the highest reported accuracy of 92.28%, followed by DenseNet121 at 91.41% and ResNet50 at 91.26%.

A WAE-augmented brain-MRI dataset is evaluated with hybrid and ensemble models, with CVMR-ViT reaching approximately 97% accuracy in the final setting.

Accurate detection of brain tumors from MRI images remains a critical challenge due to data imbalance and tumor variability. This research evaluates advanced hybrid and ensemble deep-learning models for binary brain-tumor classification. A publicly available dataset containing 253 images was expanded to 3,720 images using traditional augmentation and to 5,168 images through Wasserstein Autoencoder-based synthetic generation. The study introduces hybrid architectures including RSN50-ViTB and ensemble models including CVMR-ViT, integrating CNN, VGG, MobileNet, ResNet, and Vision Transformer architectures. Models were evaluated across original, simple-augmentation, combined, and WAE-augmentation datasets. CVMR-ViT achieved 95.0% accuracy on the original dataset, 95.92% with simple augmentation, 97.30% on the combined dataset, and approximately 97.0% with WAE-generated samples, alongside the reported precision, recall, and F1 measures.

SIGNAL joins graph neural networks with confidence-aware KNN inference to attribute synthetic speech and detect unseen generators.

We propose a unified framework for not only attributing synthetic speech to its source but also for detecting speech generated by synthesizers that were not encountered during training. This requires methods that move beyond simple detection to support both detailed forensic analysis and open-set generalization. To address this, we introduce SIGNAL, a hybrid framework that combines speech foundation models (SFMs) with graph-based modelling and open-set-aware inference. Our framework integrates Graph Neural Networks (GNNs) and a k-Nearest Neighbor (KNN) classifier, allowing it to capture meaningful relationships between utterances and recognize speech that does not belong to any known generator. It constructs a query-conditioned graph over generator class prototypes, enabling the GNN to reason over relationships among candidate generators, while the KNN branch supports open-set detection via confidence-based thresholding. We evaluate SIGNAL using the DiffSSD dataset, which offers a diverse mix of real speech and synthetic audio from open-source and commercial diffusion-based text-to-speech systems. We also test on the SingFake benchmark. Results show that SIGNAL consistently improves performance across both tasks, with Mamba-based embeddings delivering especially strong results. To the best of our knowledge, this is the first study to unify graph-based learning and open-set detection for tracing synthetic speech back to its origin.

RHYME fuses speech-foundation-model embeddings in hyperbolic and spherical spaces to improve cross-paradigm audio deepfake detection.

In this work, we address the challenge of generalizable audio deepfake detection (ADD) across diverse speech synthesis paradigms—including conventional text-to-speech (TTS) systems and modern diffusion or flow-matching (FM) based generators. Prior work has mostly targeted individual synthesis families and often fails to generalize across paradigms due to overfitting to generation-specific artifacts. We hypothesize that synthetic speech, irrespective of its generative origin, leaves behind shared structural distortions in the embedding space that can be aligned through geometry-aware modeling. To this end, we propose RHYME, a unified detection framework that fuses utterance-level embeddings from diverse pretrained speech encoders using non-Euclidean projections. RHYME maps representations into hyperbolic and spherical manifolds—where hyperbolic geometry excels at modeling hierarchical generator families, and spherical projections capture angular, energy-invariant cues such as periodic vocoder artifacts. The fused representation is obtained via Riemannian barycentric averaging, enabling synthesis-invariant alignment. RHYME outperforms individual PTMs and homogeneous fusion baselines, achieving top performance and setting new state-of-the-art in cross-paradigm ADD.

MiCuNet uses mixed-curvature fusion and temporal gating to trace emotion, manipulation, and generator source across English and Chinese synthetic speech.

In this work, we address the problem of fine-grained traceback of emotional and manipulation characteristics from synthetically manipulated speech. We hypothesize that combining semantic-prosodic cues captured by Speech Foundation Models (SFMs) with fine-grained spectral dynamics from auditory representations can enable more precise tracing of both emotion and manipulation source. To validate this, we introduce MiCuNet, a multitask framework for fine-grained tracing of emotional and manipulation attributes in synthetically generated speech. The approach integrates SFM embeddings with spectrogram-based auditory features through a mixed-curvature projection mechanism spanning Hyperbolic, Euclidean, and Spherical spaces, guided by learnable temporal gating. It simultaneously predicts original emotions, manipulated emotions, and manipulation sources on the EmoFake dataset across English and Chinese subsets. MiCuNet yields consistent improvements over conventional fusion strategies.

Fourteen ML/DL models, augmentation, and metaheuristic search predict experimental and KKL thermal conductivity from a 278-sample dataset.

The precise forecasting of thermal conductivity in nanofluids is essential for enhancement of thermal management systems within industrial contexts. This research establishes a framework that combines machine learning, deep learning, advanced augmentation, and optimization to predict experimental thermal conductivity (Exp-TC) and effective thermal conductivity based on the Koo-Kleinstreuer-Li model (KKL-TC). The dataset contains 278 samples with nanoparticle materials, base fluids, particle size, temperature, volume fraction, and thermal conductivity parameters. Polynomial and Fourier expansion-inspired augmentation and conditional variational autoencoders improve data diversity. Fourteen ML and DL models are evaluated as standalone and stacked ensembles, with Grey Wolf and Particle Swarm Optimization used for hyperparameters. CatBoost delivers the best Exp-TC performance (R2 = 0.99964, RMSE = 0.00464) and KKL-TC performance (R2 = 0.99782, RMSE = 0.00391).

The study combines augmentation, fourteen ML/DL models, and Grey Wolf/Particle Swarm optimization to predict hybrid-nanofluid density from 436 samples.

This research presents a robust and comprehensive framework for predicting the density of hybrid nanofluids using state-of-the-art machine learning and deep learning techniques. The study used a curated dataset of 436 samples and nine input parameters, with preprocessing through outlier removal and augmentation using autoencoder-based or Gaussian-noise injection. Fourteen predictive models were analyzed, with Grey Wolf Optimization and Particle Swarm Optimization used for hyperparameter search. Autoencoder-based augmentation combined with hyperparameter optimization improved predictive accuracy across models. Gradient Boosting achieved an R2 score of 0.99999 and MSE of 0.00091 among machine-learning models, while a stacked RNN with Linear Regression achieved an R2 of 0.9999, MSE of 0.0014, and MAE of 0.012 among deep-learning models.

Twelve ML/DL models, ensemble methods, augmentation, and metaheuristic tuning are evaluated on 1,269 experimental nanofluid samples.

Accurately predicting the specific heat capacity of nanofluids is critical for optimizing their performance in engineering and industrial applications. This study explores twelve machine learning and deep learning models using conventional and stacking ensemble techniques, with linear regression as a meta-learner. Particle Swarm Optimization and Grey Wolf Optimization fine-tune machine-learning hyperparameters. The research uses 1,269 experimental nanofluid samples, with inputs including nanofluid type, temperature, and volume concentration. Polynomial/Fourier expansion-inspired and autoencoder-based augmentation strategies improve generalization. The stacked multi-layer perceptron with linear regression achieves the highest predictive accuracy, with an R2 score of 0.99927, MSE of 466.06, and RMSE of 21.58. CatBoost is the best standalone machine-learning model, with an R2 score of 0.99923, MSE of 487.71, and RMSE of 22.08. Grey Wolf Optimization reduces the LightGBM MSE from 29386.43 to 6549.006.

An ensemble of CNNs and Vision Transformers classifies seven soil types and feeds an adaptive fuzzy system for crop recommendations.

This study introduces an advanced Artificial Intelligence (AI) framework for soil classification and crop recommendation, combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) in an ensemble approach, alongside an adaptive fuzzy logic-based decision system for crop suggestions. While existing research typically addresses soil classification or crop recommendation in isolation, this work integrates cutting-edge deep learning models and fuzzy logic to enhance both tasks. The methodology is divided into two phases: Phase 1 covers data collection, preprocessing, and augmentation using Cycle Generative Adversarial Networks (CycleGAN) to expand the curated dataset of 1189 soil images to 8,413, while Phase 2 focuses on training CNN and ViT models, ensembling these models, and developing a fuzzy logic system that considers soil type, nutrients, potential of hydrogen (pH), and climatic conditions for crop recommendations. Experimental results indicate models achieve classification accuracies of up to 89.32% on the original dataset, improving to 91.01% with augmented data. On the CycleGAN-augmented (CyAUG) dataset, EfficientNet v2 Large and ViT-Large/16 attain accuracies of 99.60% and 99.73%, respectively. Furthermore, an ensemble of these architectures achieves a perfect accuracy of 100%. The results are also validated by K-fold cross-validation. The research also presents 'Agro Companion,' an AI-powered tool that assists farmers in soil identification and crop selection based on geological and environmental data. This framework addresses key agricultural challenges in India, offering a high-accuracy, practical solution for improving soil classification and crop recommendation.

Herbify standardizes 6,104 images across 91 species and combines EfficientNetV2-Large with ViT-Large/16 for high-precision identification.

Herbs have historically been central to medicinal practices, representing one of the earliest forms of therapeutic intervention. While synthetic drugs are often highly effective for acute conditions, their use is frequently accompanied by adverse side effects. In response to this need, the current study introduces a computer vision framework for accurate herb identification. A novel dataset, Herbify, was compiled from two different herb datasets and refined through rigorous cleaning, preprocessing, and quality control procedures. The resulting dataset underwent standardization via the Preprocessing Algorithm for Herb Detection (PAHD), producing a refined dataset of 6104 images, representing 91 distinct herb species, with an average of about 67 images per species. Utilizing transfer learning, the research harnessed pre-trained Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), then integrated these models into an ensemble framework. EfficientNet v2-Large achieved an F1-score of 99.13%, while the ensemble of EfficientNet v2-Large and ViT-Large/16, termed EfficientL-ViTL, attained an F1-score of 99.56%. The research also introduces the Herbify application, an AI-driven framework designed to identify herbs using the developed model and return scientific names and resemblance probabilities.

A 950-image, 19-rock dataset is evaluated with transfer-learned CNN families, reaching above 99% in the preprocessed and cross-validated settings.

This study aims to tackle the obstacles linked with geological image segmentation by employing sophisticated deep learning techniques. Geological formations, characterized by diverse forms, sizes, textures, and colors, present a complex landscape for traditional image processing techniques. Drawing inspiration from recent advancements in image segmentation, particularly in medical imaging and object recognition, this research proposed a comprehensive methodology tailored to the specific requirements of geological image datasets. To establish the dataset, a minimum of 50 images per rock type was deemed essential, with the majority captured at the University of Las Palmas of Gran Canaria and during a field expedition to La Isla de La Palma, Spain. This dual-source approach ensures diversity in geological formations, enriching the dataset with a comprehensive range of visual characteristics. The study involves the identification of 19 distinct rock types, each documented with 50 samples, resulting in a comprehensive database containing 950 images. The methodology involves two crucial phases: initial preprocessing, followed by transfer learning and fine-tuning of ResNet, Inception V3, DenseNet, MobileNet V3, and EfficientNet V2 large models. DenseNet201 and InceptionV3 attained the highest accuracy of 98.49% on the original dataset; in five-fold cross-validation, MobileNet V3 large reached 99.15% accuracy. The authors report faster convergence without overfitting after preprocessing and minimal misclassifications among specific classes.

Pretrained CNNs are tuned and cross-validated for mango leaf disease classification, with Inception V3 leading the reported results.

This paper presents a comprehensive investigation into mathematical modelling and statistical analysis of mango leaf disease classification using deep learning models. The methodology includes preprocessing, hyperparameter tuning, and fine-tuning of pretrained Inception V3, MobileNet V3 Small, MobileNet V3 Large, and ResNet50 models. The reported results show 100% accuracy on validation and test datasets during initial training, while k-fold cross-validation affirms consistency and identifies Inception V3 as the leading model. Training and loss graphs and confusion matrices provide further analysis, including remaining distinctions among healthy leaf, Gall Midge, and Anthracnose classes.