adverSCarial: assessing the vulnerability of single-cell RNA-sequencing classifiers to adversarial attacks

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Machine learning (ML) algorithms are crucial for classifying cell types from single-cell RNA sequencing (scRNA-seq) data in biomedical research.
• Concerns exist regarding the vulnerability of these ML classifiers to adversarial attacks, which can manipulate outputs through crafted inputs.

Purpose of the Study:

• To introduce adverSCarial, an R package for simulating adversarial attacks on scRNA-seq data.
• To assess the robustness of scRNA-seq classifiers against various attack modes.
• To guide the development of more reliable and interpretable ML models for biomedical applications.

Main Methods:

• Development of the adverSCarial R package for simulating adversarial attacks.
• Simulation of diverse attack strategies, from subtle to aggressive modifications.
• Assessment of the vulnerability of five distinct scRNA-seq classifiers across four datasets.

Main Results:

• adverSCarial enables the simulation of a range of adversarial attacks on single-cell transcriptomic data.
• The study successfully assessed the robustness of various scRNA-seq classifiers to gene expression variations.
• Analysis revealed classifier sensitivities, providing insights for improving model reliability.

Conclusions:

• The adverSCarial package facilitates the evaluation of ML classifier security in scRNA-seq analysis.
• Understanding classifier vulnerabilities is key to developing robust models for clinical applications.
• This work contributes to enhancing the trustworthiness of ML in precision medicine.