Currently reading papers about diffusion models, their foundations and theis applications.

This project aim to find natural adversarial examples to test the reliability of human trajectory predictors using density estimation techniques. First, we conducted a litrature review on density estimation techniques, such as Masked autoregressive flow, RealNVP and Masked autoencoder for distribution estimation. We proceed by using Masked autoregressive flow(MAF) to find natural adversarial examples to test the reliability of human trajectory predictors. Subsequently, we adversarially trained LSTM based predictors and reduced the collision rate up to 35% in the case of adversarial attack on test data.

In medical analysis, supervised and labeled data is used in many cases. However, labeling medical images is extremely difficult, expensive, and time-consuming. In this project, we propose using genralized self-supervised frameworks to extract features from unlabeled images. We pre-trained a U-net encoder with SimCLR, MoCo, and SimSiam. We managed to improved IoU score after fine-tuning with annotated ones up to 6%.

I worked on ML based forgery recognition through handwriting style recognition. The aim of this project was to develop a plagiarism detection method which uses variations in writing style to identify potentially plagiarized passages. We extracted lexical, structural, and syntax features, proposed a regression model to fuse features and predict writing style, and finally, implemented an outlier detection model to find possible plagiarised segments.

I’m working as a scientific collaborator in the Image Processing Lab (IPL) at Sharif University under the supervision of Prof. Shohreh Kasaei. I am investigating Adversarial attacks against Deep Neural Networks, specifically focusing on 3D pointCloud networks (PointNet and PointNet++).