profile.jpg
Sohonsagar Singha —
turning binaries into signal, and signal into defense.
I research deep learning methods for malware detection — converting raw executables into visual and statistical representations that models can actually learn from.
Trained on threats, tuned for research.
I'm a PhD Computer Science scholar at NIT Meghalaya, working at the intersection of deep learning, machine learning and cybersecurity. My thesis explores static malware analysis from two angles: classifying malware by rendering binaries as images for CNNs to interpret, and studying how well detection models generalize across completely different malware datasets.
I'm broadly interested in computer vision and applied machine learning for security — the kind of problems where a model has to hold up outside the dataset it was trained on, not just inside it.
Selected research & builds.
A compilation of my research projects and experimental builds.
Image-Based Malware Classification
Two fusion architectures — a Dual-Base VGG16 + DenseNet121 model, and DLSE-Net, which adds LBP texture features and Squeeze-and-Excitation attention — for classifying malware rendered as grayscale images.
Custom CNN Classifier
A 10-layer CNN designed especially for malware images. Yet to be tested on real world images of cats and dogs.
Vesuvius-Challenge
This repository/notebook implements a lightweight 3D UNet pipeline to segment papyrus sheet surfaces from CT volume crops (Vesuvius Challenge). The notebook includes data loading, EDA, a compact UNet3D model, training, sliding-window inference, metric computation (Dice, Surface Dice, TopoScore approximation, VOI), and a no-retrain inference fix (use probability thresholding rather than hard argmax). The challenge in this competition is to detect and segment layers of papyrus in 3D imagery.
Object Detection
A basic YOLO11 model. It is using the Yolo11n, which is faster and somewhat accurate. Other powerful models can be used which will require more gpu power and memory, such as Yolo11x. Any image can be given to the model for prediction, you can put your images inside the "images" folder. Moreover it can also detect using webcam.
Let's talk research, PhD leads, or interesting problems.
Reach out about collaborations, opportunities, or anything malware-and-ML shaped. I usually reply within a couple of days.