Skip to main content

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.

Heron AI Security provides a reproducible research framework for studying how backdoor attacks inherited from pre-trained models behave when subjected to progressively stronger clean fine-tuning interventions. Rather than asking the binary question “did the backdoor survive?”, this framework measures the full degradation curve — tracking Attack Success Rate, trigger confidence, clean accuracy, and retention ratio at every stage of intervention.

Introduction

Understand the research motivation and how the framework approaches backdoor inheritance evaluation

Quickstart

Clone the repo, install dependencies, and run the baseline experiment in minutes

Concepts

Learn about backdoor inheritance, continuous degradation processes, and the evaluation methodology

API Reference

Explore the full public API across all eight source modules

​
What Makes This Framework Different

Traditional backdoor evaluation asks a single before/after question. Heron AI Security measures model behavior at every step of a progressive intervention, revealing degradation patterns that binary evaluation misses entirely. 
Backdoored Model
      │
      ▼
Progressive Clean Fine-Tuning
0% → 20% → 40% → 60% → 80% → 100%
      │
      ▼
Evaluate After Every Intervention
      │
      ▼
Continuous Degradation Curves
1

Clone and install

Clone the repository and install all Python dependencies into a virtual environment.
2

Configure your experiment

Edit configs/baseline.yaml to set your dataset, model, attack parameters, and intervention schedule.
3

Run the experiment

Execute python main.py to train the backdoored model and measure its degradation across all intervention strengths.
4

Inspect results

Find timestamped outputs in results/, including metrics.csv and publication-quality figures.

​
Key Capabilities

Modular Attack Pipeline

Pluggable attack module supporting BadNet and extensible to any dataset-transformation-based attack

Progressive Interventions

Apply clean fine-tuning at configurable ratios from 0% to 100% of the available clean dataset

Four Evaluation Metrics

Track ASR, clean accuracy, mean trigger confidence, and retention ratio simultaneously

Reproducible Experiments

Fixed seeds, YAML configs, timestamped outputs, and saved checkpoints ensure full reproducibility

Build docs developers (and LLMs) love

Get started for freeTalk to us