Language and Tools: Python, Scikit-Learn, Tensorflow, Keras, Matplotlib, NLTK

This research incorporates designing customized Machine Learning (ML) and Deep Learning (DL) Frameworks in order to improve Opinion Mining tasks by employing natural language features such as TF-IDF, Word Embedding, Bag of Words, etc. The research focuses on cyber-bullying recognition, fake news identification, spam filtering with robust Artificial Neural Network, Hybrid LSTM-CNN architecture and ensemble of Machine Learning algorithms. So far, the research focuses only on the benchmark datasets created by checking the authenticity from different internet sources by many researchers and publicly available dataset websites.

Before extracting NLP features and applying ML and DL algorithms, preprocessing steps such as noise removal, stemming, lemmatization, etc. are performed on the raw dataset to make a clean dataset. Till now, level-based voting classifiers have been generated which have already surpassed traditional ML algorithms and ensemble techniques. Moreover, different DL models such as ANN, CNN, LSTM, Bidirectional-LSTM and the hybrid of LSTM-CNN architecture have been devised by tuning various parameters to achieve the best results in the mentioned domain.

Already, one papers related to this research has been accepted and presented in esteemed conference (ICCTSAI 2021) and two papers (Three-level voting model for Misleading Information Detection, Ensemble based Cyberbullying Detection and Improved Spam Email Filtering) have been published at two conferences (ICICV 2021, EAIT 2021, BIM 2021). This research is currently ongoing with more advanced techniques in the field of Natural Language Processing.

Language and Tools: Python, Librosa, Praat, Scikit-Learn, Tensorflow, Keras, Matplotlib

This research work on signal processing and recognition basically consists of identifying voice diseases by extracting audio features and applying ML algorithms in the most important features. Here, the impact of Dimensionality Reduction, Audio Frame Truncation and the performance of several Acoustic Features have been analyzed. Several voice disorders (Dysphonia, Laryngitis, Renkei’s Edema) were recognized by extracting Mel-Frequency Cepstral Coefficient (MFCC), Fundamental Frequency, Shimmer, Jitter and Harmonic-to-Noise Ratio (HNR), etc.

In this study, two types of dimensionality reduction technologies- Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) have been exerted on audio features extracted from the audio files of a well-established voice database namely "Saarbruecken Voice Database (SVD)". The result analysis comprises the performance of ML and DL algorithms on the reduced voice feature matrix. Moreover, the clipping of audio frames of constant frequency ".wav" files has been considered to optimize the training time of the recognition scheme.

One paper related to this research has been published in a conference (BIM 2021) and one manuscript is being prepared to submit in a Scopus indexed journal.

Language and Tools: Python, Matplotlib, Weka, Excel

The study includes categorization of the vulnerability of migrant-friendly countries for Bangladeshi Immigrants due to Covid-19 by analyzing the return causes of migrants and other categorical features using Clustering Algorithms and Probabilistic Modeling. Probabilistic likelihood score was assigned to each attribute by Bayes theorem where unsupervised K-Means++, Agglomerative and BIRCH algorithms were applied to split twelve unlabeled countries into five separate classes.

The dataset for this research was collected from Young Power in Social Action (YPSA). Since it was an unlabeled dataset and there were some missing values, we have cleaned the dataset and preprocessed it by exploring it in WEKA. To validate the number of clusters, the Silhouette Coefficient has been measured for numerous class labels.

One paper related to this research has been published at a conference (ICICT4SD 2021). This paper has also been indexed in WHO (COVID-19 Global Literature on Coronavirus Disease). There is a plan to extend the work by adding supervised classification and subjective labeling.

Language and Tools: Python, Linux OS, PyAudioAnalysis

DNA Cryptography has been utilized to provide security to audio signals in this study. The audio samples have been converted to binary signal and the DNA sequence has been also converted to a binary signal. These two binary sequences are merged together to encode the audio. The decoding step helps to get the actual signal. This work helps to protect any audio signals to be captured by intruders in the middle of a transmission.

A paper has been published at a conference (e-ISSP 2020) related to this work.

Language and Tools: Javascript, Hyperledger, Playground

This study corresponds to provide security to Forensic Information by creating IPFS and Hyperledger based Private Blockchain System to inhibit intruder intervention. Here, the insertion, deletion or update of any forensic evidence is counted as a transaction and the information is stored in the blocks connected in the chain. This security framework is effective since it can provide robust security to this essential field with better performance than public blockchain.

One paper has been published at a conference (ACMI 2021) related to this research.