SecretLab

Dataset Value Taxonomy

This project formalizes how to evaluate and compare dataset value in AI-native empirical research. It frames practical criteria that can guide data collection, curation, and downstream security-aware analysis.

LLM Package Hallucination Analysis

This work investigates software supply-chain risk created by hallucinated dependencies in code-generating LLM outputs. It combines large-scale measurement with actionable insights to reduce package-level attack surface in developer workflows.

Prompt Refinement for Image Regeneration

This research examines how people infer prompts from AI-generated art, how contextual circumstances influence inference quality, and how iterative human-driven prompt refinement affects regeneration outcomes. Through human-subject studies and controlled regeneration tasks, it identifies when prompt inference is reliable, where it fails, and how structured refinement workflows can improve quality, intent alignment, and user control.

EtherBee

EtherBee constructs a global Ethereum dataset that combines node performance measurements, honeypot interactions, and full network-session traces to support reproducible security analysis of P2P blockchain infrastructure. Building on this data, the project provides initial evidence of elevated reconnaissance activity against overlay-network nodes and characterizes attacker scanning behavior across the ecosystem.

Spiking Neural Networks in Federated Learning

This project studies the performance and efficiency trade-offs of spiking neural networks in vertical federated learning settings. It characterizes where SNN-based approaches are beneficial and where practical deployment constraints remain.

NinjaDoH

NinjaDoH develops a censorship-resistant moving-target DNS-over-HTTPS architecture using hyperscaler infrastructure and IPNS. The system is designed to improve resolver survivability and preserve access under active blocking pressure.

Project page: ninjadoh.secretlab.page