Research Experience

• Study of Spanning tree with Maximum number of leaves

  Spanning Tree of a graph is a tree where all nodes are connected. A graph might have multiple spanning tree. A maximum leaves spanning tree (MLST) is a tree whose number of leaves is maximum. Finding MLST of a graph is a NP-hard problem. It is also APX-hard, which means it's approximation ratio is bounded by a constant. By finding the MLST of a network, we can find the backbone of the network, which can be used to refuce cost of the network. In this work, we tried to find an approximate algorithm for the MLST of a graph. We provide a heuristic algorithm following bottom-up approach.

• Chakma Handwritten Characters recognition and Multipurpose Dataset

  The Chakma language is a endangered language spoken by the Chakma people from Bangladesh. Not so many resources are available in Chakma language. Here we introduce a novel multilabel dataset for the Chakma language. The dataset contains 111,382 handwritten character images. We also introduce a novel data collection method and provide a baseline model and provide statistics, comparing with the state-of-the-art models. We provide a novel pipeline for handwritten character recognition. First we convert handwritten characters to font type images using GAN, then we use a Visual Transformer based model for classification.

• Multilingual machine translation for Chakma

  In this work we introduce a novel bilingual and monolingual corpus for the Chakma language. The dataset consists of 13,000 sentences, collected from various ources including online crowd sourcing. We then expriment with NMT and SMT.

• Android malware detection

  In this work we have curated a dataset from various sources. Then reverse engineered the APKs to for source code analysing using APKtool for decompiling. We made a code parser and extracted import, method invocation and parameter based features from the source code. Afterwords, recursive feature enimilation and PCA was used for feature reduction. Then we ran tree based models for classification of benign and malicious apps. We ran feature importance on the collected dataset and found that, method invocation based featuers are the most important features.