[Paper Review] DeepRobust: A PyTorch Library for Adversarial Attacks and Defenses

Yaxin Li, Wei Jin+2 more|arXiv (Cornell University)|May 13, 2020

Adversarial Robustness in Machine Learning34 references78 citations

TL;DR

DeepRobust is a PyTorch-based library providing over 10 image attacks, 8 image defenses, 9 graph attacks, and 4 graph defenses, aimed at fostering adversarial research across image and graph domains.

ABSTRACT

DeepRobust is a PyTorch adversarial learning library which aims to build a comprehensive and easy-to-use platform to foster this research field. It currently contains more than 10 attack algorithms and 8 defense algorithms in image domain and 9 attack algorithms and 4 defense algorithms in graph domain, under a variety of deep learning architectures. In this manual, we introduce the main contents of DeepRobust with detailed instructions. The library is kept updated and can be found at https://github.com/DSE-MSU/DeepRobust.

Motivation & Objective

• Motivate the need for a unified, extensible platform to study adversarial attacks and defenses across both image and graph data.
• Provide a comprehensive set of representative attack and defense algorithms to enable systematic experimentation.
• Facilitate easy integration with different network architectures and standard datasets for rapid research iteration.

Proposed method

• Organize attack and defense implementations into two main components: image and graph packages with clearly defined base classes and subclassed algorithms.
• Implement a suite of image attacks (LBFGS, FGSM, PGD, CW, OnePixel, DeepFool, BPDA, Universal, Nattack) and defenses (adversarial training, gradient masking, detection).
• Implement graph attacks (targeted and untargeted) and defenses on a GCN backbone, with datasets such as Cora, Citeseer, Pubmed, etc.
• Describe mathematical formulations for key attacks and defenses, including constrained optimization problems and objective functions (e.g., L2 distance, cross-entropy losses, max-margin style criteria).
• Provide practical usage workflows and example snippets to train, attack, and defend models within the library.
• Ensure environment requirements and setup are documented for reproducibility.

Experimental results

Research questions

• RQ1 What attack and defense algorithms are most representative for images and graphs within a unified PyTorch library?
• RQ2 How can a single framework support both image and graph adversarial methods with consistent interfaces?
• RQ3 How effective are standard defense strategies (adversarial training, gradient masking, detection) when applied to diverse modalities?
• RQ4 Can researchers efficiently experiment across architectures and datasets using a centralized library?

Key findings

• The library currently contains more than 10 image attacks and 8 image defenses, plus 9 graph attacks and 4 graph defenses across multiple architectures.
• It provides a two-component design (image and graph) with base classes to support extensibility and uniform APIs.
• The paper outlines concrete implementations and usage patterns for training, attacking, and evaluating models within DeepRobust.
• Examples demonstrate practical workflows for training networks, generating adversarial examples, applying defenses, and evaluating robustness.
• DeepRobust emphasizes up-to-date coverage and ongoing updates to include new models and datasets.