MACHINE LEARNING SECURITY SYSTEMS_

XSS-Shield

Problem

ML-based XSS detectors are evaded by payloads hidden under several layers of encoding.

My work

First author. Built and released a public XSS dataset with code and trained models, a preprocessing pipeline that unwraps multi-layer encodings, and an XAI toolkit that highlights the executable patterns behind each decision.

Result

On the released dataset: macro-F1 0.9958 (CharCNN) and 0.9905 (Transformer), with under 2% false positives.

INSIGHT

Problem

Fraud detection on mixed numeric and string features, where fraud is rare and decisions must be explainable.

Approach

A cluster-based framework that reduces the dimensionality of string features without information loss (with proof), a cluster-based resampling strategy for the minority class, SHAP explanations and a noise-analysis module.

Result

Near-linear time complexity in dataset size (theoretical and empirical scalability analysis) and improved minority-class detection.

WalteR

Problem

New and forgotten assets are where bugs hide, and manual reconnaissance finds them too late.

My work

Independent project. Automated subdomain discovery by parsing certificate-authority Merkle trees, plus a watch tower that tracks infrastructure changes and pushes alerts through a Discord bot backed by GraphQL.

KAN from Scratch

Goal

Understand Kolmogorov-Arnold Networks by building one without a deep-learning framework.

My work

My own scalar autograd engine, learnable B-spline edge functions (de Boor) with a SiLU base, KAN layers, an SGD optimiser and Graphviz plots of the computation graph.

Scope

Checked on a four-point XOR-style regression; no MLP baseline comparison yet. Neural Networks course, 2024.

MNIST GAN

My work

A generator (64-d noise to 784 pixels) and a discriminator, each with two 256-unit layers, trained with BCE loss and Adam for 300 epochs on 70,000 MNIST digits.

Scope

Results are judged visually from sample grids; no quantitative metric such as FID. Neural Networks course.

Raft Consensus

My work

Leader election and log replication in Python, with gRPC between nodes, MongoDB log storage and a three-node Docker Compose cluster.

Testing

An integration test checks log replication across the three nodes.

Scope

Educational implementation for the Operating Systems course. Separate from the production Raft work at NexaLoak.

Geometric A* Maze Solver

Maze solving with a geometric A* heuristic. Artificial Intelligence course.

Cryptor

An automated trading bot that acts on traders' signals, integrated with several exchanges. Independent project.