BackdoorBench

A visual guide to the comprehensive benchmark for backdoor attacks and defenses in deep learning.

16

Attack Methods

28

Defense Methods

8

Model Architectures

4

Supported Datasets

The BackdoorBench Ecosystem

Attacks vs. Defenses

The benchmark places a strong emphasis on defenses, providing a robust suite of tools to counter emerging threats.

Supported Model Architectures

A diverse range of modern neural network architectures are supported for comprehensive evaluation.

Getting Started: A Simple Workflow

Install & Setup

Clone the repo and configure the environment.

Launch Attack

Run an attack script to generate a backdoored model.

Apply Defense

Use a defense method on the compromised model.

Analyze Results

Evaluate performance with built-in analysis tools.

In-Depth Method Analysis

Attack Method Categories

Attacks are categorized by their approach, from simple poisoning to complex, input-aware triggers.

✨ Explain an Attack

Defense Method Categories

Defenses range from model patching and pruning to runtime detection and data sanitization.

✨ Suggest Defenses

Powerful Analysis Toolkit

Beyond attacks and defenses, BackdoorBench provides a rich set of tools for model and data analysis to understand vulnerabilities deeply.

📊

T-SNE / UMAP

🧠

Neuron Activation

🔥

Grad-CAM

📉

Loss Landscape

🕸️

Network Structure

🔢

Shapely Value

🔬

Feature Map

📈

Hessian Eigenvalues

Metrics Evaluation

🔀

Confusion Matrix