Selected work

Representative Papers

Enhancing Plagiarism Detection in Marathi with a Weighted Ensemble of TF-IDF and BERT Embeddings for Low-Resource Language Processing

Jan 09, 2025

To address the low accuracy of plagiarism detection in the low-resource language Marathi, this paper proposes a weighted ensemble method integrating statistical and semantic features. Specifically, it is the first to combine multilingual BERT sentence embeddings—fine-tuned to capture deep semantic representations—with TF-IDF–based lexical features, and jointly leverages SVM and Random Forest classifiers. A dynamic weighted voting mechanism is further designed to balance lexical, syntactic, and semantic signals. Evaluated on a newly constructed Marathi plagiarism dataset, the system achieves an F1-score of 0.89—outperforming the best single model by 7.2% and significantly surpassing conventional approaches. This work empirically validates the effectiveness of semantic–statistical fusion for plagiarism detection in low-resource languages and establishes a reusable technical framework for NLP applications targeting under-resourced languages.

1 citationsRead paper

CAST: Cross-Attentive Spatio-Temporal feature fusion for Deepfake detection

Jun 26, 2025

Existing methods typically model spatial and temporal features separately and fuse them superficially (e.g., via averaging or concatenation), limiting their ability to capture fine-grained, time-varying artifacts—such as eye blinking or lip distortion—thereby constraining deepfake video detection performance. To address this, we propose an end-to-end cross-modal cross-attention mechanism that enables temporal features to dynamically attend to salient spatial regions, achieving tightly coupled spatiotemporal modeling. Our architecture jointly leverages CNNs for spatial artifact extraction and Transformers for temporal inconsistency modeling, augmented by a bidirectional cross-attention module for unified optimization. Evaluated on FaceForensics++, our method achieves 99.49% AUC; under cross-dataset generalization, it attains 93.31% AUC—demonstrating substantial improvements in fine-grained artifact localization accuracy and model robustness.

0 citationsRead paper

Efficient Malware Detection with Optimized Learning on High-Dimensional Features

Jun 18, 2025

To address the high computational cost and poor generalizability of malware detection caused by high-dimensional binary features (2,381 dimensions), this paper proposes a dual-path dimensionality reduction framework integrating XGBoost-based feature selection and Principal Component Analysis (PCA). Evaluated systematically on a unified multi-source dataset (EMBER-2018, ERMDS, BODMAS), the framework explores lightweight configurations ranging from 128 to 384 dimensions. It is the first work to empirically validate the optimal synergy between XGBoost-selected features and LightGBM modeling: achieving 97.52% accuracy using only 384 dimensions (5.4% of the original), with training time of 61 minutes and memory consumption of 30 GB. Cross-domain evaluation on TRITIUM and INFERNO yields 95.31% and 93.98% accuracy, respectively—demonstrating substantial improvements in both efficiency and generalization capability.

0 citationsRead paper

Hybrid Deep Learning Model for Multiple Cache Side Channel Attacks Detection: A Comparative Analysis

Jan 28, 2025

This study addresses the challenge of detecting remote cache-side-channel fingerprinting attacks—such as cryptographic key and user behavior inference—in multi-tenant environments. We propose a hybrid deep learning method that jointly leverages feature extraction and temporal modeling capabilities. Specifically, we design a novel CNN-LSTM multimodal collaborative architecture, the first to enable unified detection across diverse cache-side-channel attack types. To capture cache access temporal characteristics, we introduce a trace-specific feature enhancement strategy incorporating sliding-window preprocessing and trace normalization. Evaluated on real-world cache trace datasets, our model achieves a detection rate of 99.96% and a false positive rate below 0.02%, outperforming five baseline models (MLP, CNN, RNN, LSTM, GRU) by an average of 1.7 percentage points. The approach significantly improves the identification of covert cache-based interactions between co-resident processes.

0 citationsRead paper

BERTopic for Topic Modeling of Hindi Short Texts: A Comparative Study

Jan 07, 2025

This study addresses topic modeling for Hindi short texts—a low-resource, high-ambiguity setting—by systematically benchmarking BERTopic against eight mainstream approaches: LDA, NMF, LSI, ARTM, PLSA, ETM, CTM, and Top2Vec. To enhance semantic representation of short texts, we employ six contextual embedding models (e.g., mBERT, IndicBERT) and propose a multi-embedding adaptation strategy. Experimental results demonstrate that BERTopic consistently achieves significantly higher coherence scores across all topic numbers, providing the first empirical validation of its robustness and superiority for Hindi short-text topic modeling. Our key contributions are: (1) the first comprehensive, systematic benchmark study dedicated to Hindi short-text topic modeling; and (2) a paradigm shift from bag-of-words assumptions to embedding-driven topic discovery, effectively overcoming inherent limitations of traditional count-based models in sparse, ambiguous short-text domains.

0 citationsRead paper

Recent publications

Latest Papers

CAST: Cross-Attentive Spatio-Temporal feature fusion for Deepfake detection

Jun 26, 2025

Existing methods typically model spatial and temporal features separately and fuse them superficially (e.g., via averaging or concatenation), limiting their ability to capture fine-grained, time-varying artifacts—such as eye blinking or lip distortion—thereby constraining deepfake video detection performance. To address this, we propose an end-to-end cross-modal cross-attention mechanism that enables temporal features to dynamically attend to salient spatial regions, achieving tightly coupled spatiotemporal modeling. Our architecture jointly leverages CNNs for spatial artifact extraction and Transformers for temporal inconsistency modeling, augmented by a bidirectional cross-attention module for unified optimization. Evaluated on FaceForensics++, our method achieves 99.49% AUC; under cross-dataset generalization, it attains 93.31% AUC—demonstrating substantial improvements in fine-grained artifact localization accuracy and model robustness.

0 citationsRead paper

Efficient Malware Detection with Optimized Learning on High-Dimensional Features

Jun 18, 2025

To address the high computational cost and poor generalizability of malware detection caused by high-dimensional binary features (2,381 dimensions), this paper proposes a dual-path dimensionality reduction framework integrating XGBoost-based feature selection and Principal Component Analysis (PCA). Evaluated systematically on a unified multi-source dataset (EMBER-2018, ERMDS, BODMAS), the framework explores lightweight configurations ranging from 128 to 384 dimensions. It is the first work to empirically validate the optimal synergy between XGBoost-selected features and LightGBM modeling: achieving 97.52% accuracy using only 384 dimensions (5.4% of the original), with training time of 61 minutes and memory consumption of 30 GB. Cross-domain evaluation on TRITIUM and INFERNO yields 95.31% and 93.98% accuracy, respectively—demonstrating substantial improvements in both efficiency and generalization capability.

0 citationsRead paper

Hybrid Deep Learning Model for Multiple Cache Side Channel Attacks Detection: A Comparative Analysis

Jan 28, 2025

This study addresses the challenge of detecting remote cache-side-channel fingerprinting attacks—such as cryptographic key and user behavior inference—in multi-tenant environments. We propose a hybrid deep learning method that jointly leverages feature extraction and temporal modeling capabilities. Specifically, we design a novel CNN-LSTM multimodal collaborative architecture, the first to enable unified detection across diverse cache-side-channel attack types. To capture cache access temporal characteristics, we introduce a trace-specific feature enhancement strategy incorporating sliding-window preprocessing and trace normalization. Evaluated on real-world cache trace datasets, our model achieves a detection rate of 99.96% and a false positive rate below 0.02%, outperforming five baseline models (MLP, CNN, RNN, LSTM, GRU) by an average of 1.7 percentage points. The approach significantly improves the identification of covert cache-based interactions between co-resident processes.

0 citationsRead paper

Enhancing Plagiarism Detection in Marathi with a Weighted Ensemble of TF-IDF and BERT Embeddings for Low-Resource Language Processing

Jan 09, 2025

To address the low accuracy of plagiarism detection in the low-resource language Marathi, this paper proposes a weighted ensemble method integrating statistical and semantic features. Specifically, it is the first to combine multilingual BERT sentence embeddings—fine-tuned to capture deep semantic representations—with TF-IDF–based lexical features, and jointly leverages SVM and Random Forest classifiers. A dynamic weighted voting mechanism is further designed to balance lexical, syntactic, and semantic signals. Evaluated on a newly constructed Marathi plagiarism dataset, the system achieves an F1-score of 0.89—outperforming the best single model by 7.2% and significantly surpassing conventional approaches. This work empirically validates the effectiveness of semantic–statistical fusion for plagiarism detection in low-resource languages and establishes a reusable technical framework for NLP applications targeting under-resourced languages.

1 citationsRead paper

BERTopic for Topic Modeling of Hindi Short Texts: A Comparative Study

Jan 07, 2025

This study addresses topic modeling for Hindi short texts—a low-resource, high-ambiguity setting—by systematically benchmarking BERTopic against eight mainstream approaches: LDA, NMF, LSI, ARTM, PLSA, ETM, CTM, and Top2Vec. To enhance semantic representation of short texts, we employ six contextual embedding models (e.g., mBERT, IndicBERT) and propose a multi-embedding adaptation strategy. Experimental results demonstrate that BERTopic consistently achieves significantly higher coherence scores across all topic numbers, providing the first empirical validation of its robustness and superiority for Hindi short-text topic modeling. Our key contributions are: (1) the first comprehensive, systematic benchmark study dedicated to Hindi short-text topic modeling; and (2) a paradigm shift from bag-of-words assumptions to embedding-driven topic discovery, effectively overcoming inherent limitations of traditional count-based models in sparse, ambiguous short-text domains.

0 citationsRead paper