Minhajul Islam

Faysal Mahmud, Yusha Abdullah, Minhajul Islam et al.

Deepfake technology is widely used, which has led to serious worries about the authenticity of digital media, making the need for trustworthy deepfake face recognition techniques more urgent than ever. This study employs a resource-effective and transparent cost-sensitive deep learning method to effectively detect deepfake faces in videos. To create a reliable deepfake detection system, four pre-trained Convolutional Neural Network (CNN) models: XceptionNet, InceptionResNetV2, EfficientNetV2S, and EfficientNetV2M were used. FaceForensics++ and CelebDf-V2 as benchmark datasets were used to assess the performance of our method. To efficiently process video data, key frame extraction was used as a feature extraction technique. Our main contribution is to show the models adaptability and effectiveness in correctly identifying deepfake faces in videos. Furthermore, a cost-sensitive neural network method was applied to solve the dataset imbalance issue that arises frequently in deepfake detection. The XceptionNet model on the CelebDf-V2 dataset gave the proposed methodology a 98% accuracy, which was the highest possible whereas, the InceptionResNetV2 model, achieves an accuracy of 94% on the FaceForensics++ dataset. Source Code: https://github.com/Faysal-MD/Unmasking-Deepfake-Faces-from-Videos-An-Explainable-Cost-Sensitive-Deep-Learning-Approach-IEEE2023

Aurprita Mahmood, Sabrin alam, Neloy kumer Sagor et al.

Mathematical Word Problems (MWPs) are among the most challenging tasks in natural language processing because they require both linguistic understanding and multi-step numerical reasoning. While Chain-of-Thought (CoT) prompting has shown promise, its linear structure often propagates errors, limiting overall effectiveness. To address this limitation, we present the a systematic study of Tree-of-Thought (ToT) reasoning for Bengali MWPs using the SOMADHAN dataset. Owing to computational and token-cost constraints, we evaluate a curated set of 100 representative problems across multiple large language models (LLMs), including GPT-OSS and LLaMA variants, under standard prompting, CoT, and ToT strategies. Our results show that CoT improves baseline accuracy from 78% (standard prompting) to 83% on average, while ToT further increases performance by up to 5 percentage points, achieving 88% accuracy with GPT-OSS-120B. These improvements highlight that ToT is particularly effective in medium-to-large-scale models but may offer less advantage for smaller ones. Overall, our findings establish ToT as a robust framework for solving mathematical problems in low-resource languages such as Bengali. More broadly, this study shows that structured reasoning methods like ToT can provide more reliable and globally consistent outcomes than CoT, paving the way for better reasoning strategies in multilingual NLP.