Heron AI Security is a research framework that measures how inherited backdoor failures in NLP models evolve throughout progressive clean fine-tuning — tracking four complementary metrics at every intervention stage instead of recording a single pass/fail result at the end. By treating backdoor persistence as a continuous degradation process rather than a binary event, the framework reveals rich behavioural patterns that a before-and-after comparison would discard entirely: how steeply attack success rate falls, how quickly trigger confidence collapses, and how well clean accuracy is preserved throughout the intervention. Traditional backdoor evaluation asks a single binary question — does the backdoor survive fine-tuning? — which is useful but fundamentally coarse. An inherited backdoor that drops from 95 % attack success rate to 5 % over five progressive intervention stages tells a fundamentally different story than one that stays at 95 % through four stages and then collapses. Binary evaluation cannot distinguish these two trajectories. Heron AI Security was built to make that distinction tractable and reproducible.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/Antisource/heronaisec/llms.txt
Use this file to discover all available pages before exploring further.
Research Question
Can inherited model failures be evaluated more effectively as continuous degradation processes rather than binary survival outcomes?The framework operationalises this question by measuring model behaviour throughout progressively stronger interventions — 0 %, 20 %, 40 %, 60 %, 80 %, and 100 % of the available clean training data — and recording Attack Success Rate, Clean Accuracy, Mean Trigger Confidence, and Retention Ratio at every stage.
Scientific Contribution
The project contributes a new evaluation methodology, not a new attack or defence. Understanding this boundary is essential before deciding whether the framework fits your research workflow.What this framework contributes
- ✅ A methodology for evaluating inherited model failures as continuous degradation processes
- ✅ A reproducible evaluation framework for comparing intervention strategies
- ✅ A modular research pipeline designed to support future extensions
- ❌ A new backdoor attack
- ❌ A new defense mechanism
- ❌ A new optimization algorithm
Architecture Overview
The repository is organised around scientific variables rather than software implementation details. Each layer of the pipeline maps directly to an independent research dimension — attack, model, intervention, and evaluation methodology can all evolve without touching the others.cf at 15 % poison rate, and progressive clean fine-tuning as the intervention.
Key Features
Quickstart
Clone the repository, install dependencies, and run the baseline experiment in under ten minutes.
Backdoor Inheritance
Learn what backdoor inheritance is, why it matters, and how continuous evaluation changes what you can observe.
Running Experiments
Configure and execute experiments using the YAML-driven pipeline, from prototype subsets to full-scale runs.
Attacks API
Explore the attacks module, understand how BadNet poisoning is applied, and learn how to extend the pipeline with new attack families.
Tech Stack
The framework is built on standard scientific Python tooling and the Hugging Face ecosystem. No proprietary dependencies are required.| Layer | Library |
|---|---|
| Language | Python 3.8+ |
| Deep Learning | PyTorch |
| Model Hub & Tokenisation | Hugging Face Transformers |
| Dataset Loading | Hugging Face Datasets |
| Numerical Computing | NumPy |
| Tabular Results | Pandas |
| Figures | Matplotlib |