Skip to content

overview.md

Latest commit

 

History

History
58 lines (46 loc) · 2.58 KB

overview.md

File metadata and controls

58 lines (46 loc) · 2.58 KB
Raw

Architecture Overview

Mithridatium is a Python package and CLI for running backdoor-detection defenses against image classification models. The current CLI runs one selected defense per command.

flowchart TD
    A[User runs mithridatium detect] --> B[Parse CLI options]
    B --> C[Load model]
    C --> D[Build preprocessing config and dataloader]
    D --> E[Check defense compatibility]
    E --> F[Run selected defense]
    F --> G[Build JSON report]
    G --> H[Validate report schema]
    H --> I[Write report or stdout]
Loading

Main Components

Area Important files Purpose
CLI mithridatium/cli.py Typer commands, options, defense dispatch, report writing
Model loading mithridatium/loader.py, mithridatium/loader_hf.py Local checkpoint loading and Hugging Face model wrapping
Preprocessing mithridatium/utils.py Dataset configs, dataloaders, normalization, image sizes
Defenses mithridatium/defenses/ FreeEagle, STRIP, MMBD, and AEVA implementations
Reporting mithridatium/report.py, reports/report_schema.json Report payloads, summaries, JSON schema validation
UI/service app.py, mithridatium/gradio_app.py, mithridatium/service.py Streamlit app, Gradio app, and service-oriented wrappers
Tests tests/ Unit and integration tests for loaders, reports, attacks, and defenses

Current Detection Model

The CLI supports these defenses individually:

  • freeeagle
  • strip
  • mmbd
  • aeva

The current system does not combine all defenses in one command. Users choose one defense with --defense.

Important Design Notes

  • Local models are expected to be PyTorch .pt or .pth checkpoints.
  • Local ResNet checkpoints are auto-detected as standard ResNet-18 or CIFAR-style ResNet-18 based on conv1.weight.
  • Hugging Face support wraps AutoModelForImageClassification models as plain PyTorch classifiers.
  • Hugging Face compatibility depends on model architecture, processor metadata, and preprocessing alignment.
  • FreeEagle is white-box and currently ResNet-family only.
  • STRIP and AEVA need representative input data.
  • Dataset mismatch can change the behavior of data-dependent defenses.
  • A hosted Streamlit demo is available at https://huggingface.co/spaces/williamphoenix/Mithridatium.
  • The local Streamlit entry point is app.py; the CLI also exposes a Gradio UI through mithridatium ui.

Related Docs