DKDhanashree S. Kulkarni

Dhanashree S. Kulkarni

*This is an algorithmically generated author record based on publicly available information.

Dhanashree S. Kulkarni has a total of 9 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Real-Time Deepfake Detection via Frame-Level EfficientNet Ensemble and Client-Server Deployment

    by Vishwakalyan Patil, Akshay Sarapure, Harshal Poriwade, Jyoti Kamble, Sonam Bhandurge, Dhanashree Kulkarni, Anand Deshpande

    DeepFake technology threatens the integrity of digital content by allowing the generation of hyperrealistic but misleading content. In this paper, we propose a frame-based DeepFake detection framework that aims to enhance detection accuracy and computational efficiency. Our approach takes advantage of three large-scale datasets—Celeb-DF, FaceForensics++ (FF++), and DeeperForensics—consisting of close to 17GB of training data. Preprocessing includes frame extraction and face alignment with Dlib’s 68-point facial landmark detection to maintain high-resolution facial features. Two CNN models, EfficientNet-B4 and EfficientNet-B5, are ensembled and trained to detect coarse and fine-grained forgery artifacts. The model performs AUC of 0.9958 and F1-score of 0.9726 on benchmark sets, on par with state-of-the-art techniques. To ensure real-world deployability, the detection pipeline is incorporated within a cross-platform client-server setup in a modular structure. A responsive web frontend and an Android app provide real-time predictions and Grad-CAM-based visualizations. This end-to-end solution bridges the gap between research and deployment by providing a scalable, interpretable, and accessible platform for DeepFake detection. Unlike previous methods, our system integrates multimodal detection in videos with a real-time, cross-platform deployment pipeline, bridging research with practical application.

  • Open Access

    Impact of Textual Length on Lexicon-Based Sentiment Classification for Hindi Reviews

    by Dhanashree S. Kulkarni, Anand B. Deshpande, Sonam A. Bhandurge, Vania V. Estrela

    Cureus Journals The performance of sentiment analysis is affected by different properties like dataset size or training corpus size, feature extraction, length of target documents, preprocessing constraints, length of the review/word count, readability, purity of data, subjectivity of data, etc. Efforts need to be taken to not only increase the performance of the algorithms but also to investigate the impact of some of the linguistic properties of the dataset on the performance. One such data property is word count. The lexicon model was experimented for effect of word count and was tested on both the single and multi-domain datasets to find the effect of word count on accuracy and other parameters by splitting into four categories, where each category or group had a different word count. The experimentation results indicated that the accuracy was significantly high for sentences that had shorter words. After a particular threshold, the accuracy starts decreasing as the number of words in the sentence goes on increasing. The statistical significance of the impact of word count on the classification accuracy clearly indicates that the impact of the review words count on the classification accuracy is statistically significant at a 99% confidence level. Less than 80 words is an ideal level for the review words count.