20 March 2025 in IEEE Access
In the digital era, the proliferation of online reviews through videos has been meteoric and driven by recent technological advancements. The sentiments expressed in these videos drive consumer reliance, decision-making and may shape perception building of general public. The sentiment classification of videos is emerging as a vital area of research in Natural Language Processing (NLP) and is challenging due to complexity of emotions and temporal aspects of modalities. These challenges become more complex for low-resourced languages such as Urdu. Addressing the need for a comprehensive study in this realm, we introduce a generic framework, Urdu Multimodal Sentiment Analysis (UMSA), for emotion detection and sentiment classification of videos. Our fusion-based approach highlights that incorporating additional modalities and cross-modal interactions significantly enhances the analysis. Our research culminates in the creation of the Urdu Sentiments Dataset (USD), a comprehensive collection of Urdu video reviews. In this study, we classify videos using a two-phase analysis approach that incorporates early fusion and ensembling. After fusion, we perform ensembling of two models for each modality: audio, text, and frames. We utilize Long Short-Term Memory (LSTM) networks and Random Forest Classifier for audio. Text-based analysis is conducted using Logistic Regression and the Bidirectional Encoder Representations from Transformers (BERT) model. For frames, we employ Random Forest and Convolutional Neural Networks (CNN). Afterwards, we implement model ensembling across the three modalities. This multimodal integration proves essential in providing a clearer and more comprehensive understanding of the sentiments conveyed and achieved more than 80% accuracy. The validation of this framework is reinforced through a comprehensive case study approach. This independent validation highlights the framework’s robustness and adaptability to real-world scenarios.
24 February 2025 in IEEE Access
Cybercrimes frequently manifest through the utilization of the internet. Spoofing attacks epitomize an advanced facet of contemporary cybercrime. Considerable research has been undertaken in the domain of these diverse spoofing attack techniques. Recent spoofing attacks predominantly concentrate on voice spoofing, driven by the vulnerabilities of Automatic Speaker Verification systems. Such type of attacks contains significant threats, resulting in fraudulent activities including imitation of someone’s voice for financial scam, bypassing security locks, or getting sensitive information and data. To address these concerns, this research focuses on the detection of voice spoofing attacks by utilizing a hybrid VGGish-LSTM model. While previous transfer learning methodologies have essentially been applied to image classification tasks, our methodology focuses on voice classification. In comparison to VGGish, another transfer learning model i.e. YAMNet is also utilized. These models support the extraction of significant features i.e. embeddings, capable of mitigating bias and facilitating analysis with limited datasets. For classification, LSTM and 1D CNN models are applied on the extracted embeddings. The evaluation uncovers that the hybrid VGGish-LSTM model exhibits better performance than alternative configurations. We achieved a remarkable accuracy of 99.11% with an equal error rate of 1.18% on ASV Spoof 2019 LA dataset. This research helps the general public to recognize spoof attacks, thereby helping in the prevention of frauds by the detection of synthesized speech.