A New U-Net based system for multi-cultural wedding image classification

Identifying where the image might have taken or the clothes wore by people may be helpful for us in various ways. Some applications can be, implementing an end to end sytem to enhance an image based on the culture of the particular place, where it can be used as a first step. At the same time, the same photos can be used for authenticating genuine marriage or fake marriage and kinship verification. Therefore here I have presented novel method based on the combination of U-Net, Convolutional Neural Network and Random Forest for classification of multicultural wedding images.

At first we segmented the region of interest, which included the portions of dresses or costumes with wedding couples. For this task we implemented UNet architecture which has ability to segment irrespective of background complexities. After segmenting region of interest, to extract distinct features for classification of images, we explore the Convolutional Neural Network (CNN) as it has the ability to extract context information between the pixels in the images. Features extracted from CNN architecture is used as an input in Random Forest (RF), for classification of images of different cultures. The reason to use RF for classification is that it does not require a large number of samples and it avoids the cause of overfitting.

The proposed CNN architecture shown in below image can be found in the paper link given below.

Hope Speech Detection on Social Media Platforms

Over the last few decades, the impact of social media on an individual has increased exponentially, which has led to certain demerits such as cyber bullying. In this project, we built a NLP model to classify social comments based on Hope Speech, Non-Hope Speech and Neutral speech ontology.

The model consists of three units, the pre-processing unit, the feature extraction and data balancing unit, and the classification unit (as shown in the figure). In the pre-processing unit, the dataset is cleaned up by removing unwanted text portions, expanding the abbreviations, and using lemmatization to convert words to their root words. The output of this step is then fed to the feature extraction unit that uses the TF-IDF vectorization technique to identify the importance of specific words or phrases in the dataset. Therefore, the sentences in the dataset are converted to vectors of different values. These vectors are then used to balance the dataset using data balancing methods like SMOTE and ADASYN by adding some synthetic samples

The output from the above steps is given to the final unit, which constitutes various traditional Machine Learning models like Naïve Bayes, Logistic Regression and SVM. These models were implemented using the scikit-learn library available in python. We experimented the result with certain traditional model and also implemented BERT architecture.

Random Decision Forest approach for Mitigating SQL Injection Attacks

A Structured Query Language injection (SQLi) attack is also commonly referred to as a code injection attack, is a security vulnerability which allows access to data which is intended to be protected and inaccessible to unauthorized personnel. Therefor to mitigate such attacks we proposed a Machine Learning model which used Random Decision Trees as classifier.

We Collected data(malicious sql queries) and identified features which were comment characters, number of semicolons, Logical operators, True conditions, keywords and wildcard characters to streamline the data for better insights. Scored a list of keywords according to their risk factors for better understanding the data. This model was then trained over the new created dataset where we implemented Random Decision Trees classifier.