adverSCarial

adverSCarial generates and analyzes the vulnerability of single-cell RNA-sequencing (scRNA-seq) classifiers to adversarial attacks, enabling robustness assessment against gene expression variations.

Key Features:

• Classifier Integration Format: Provides a versatile format for integrating any type of classifier into the adversarial analysis framework.
• Single-Gene Attack: Simulates attacks by making a small modification to the input to alter the classification output.
• Max-Change Attack: Simulates attacks by making a large modification to the input without changing its classification.
• CGD Attack: Implements an attack based on estimated gradient descent.

Scientific Applications:

• Classifier Robustness Assessment: Evaluates the robustness of scRNA-seq classifiers to adversarial attacks across multiple ML algorithm classes.
• Vulnerability Simulation: Simulates adversarial attacks ranging from expanded but undetectable modifications to aggressive and targeted ones.
• Interpretability Analysis: Exposes inner workings and sensitivities of classifiers to guide development of more reliable models for biomedical research and clinical applications.

Methodology:

The package implements single-gene attacks, max-change attacks, and CGD attacks based on estimated gradient descent.