OmicsDefense: The First Unified Framework for Defending Against Backdoor Attacks in Single-cell Foundation Models

Bioclaw Team, Peiting Shi

Keywords: single-cell foundation models, backdoor defense, AI security, bioinformatics

TL;DR: We present OmicsDefense, a unified defense framework that detects and filters backdoor attacks in single-cell transcriptomics foundation models.

Abstract: Single-cell foundation models have become indispensable tools for cellular heterogeneity analysis, disease mechanism discovery, and clinical diagnostics, yet they face critical biosecurity vulnerabilities from backdoor attacks that remain unaddressed. **We present OmicsDefense, the first unified defense framework for single-cell foundation models, featuring Weighted Divergence Filtering (WDF) for poisoned sample detection.** We apply our framework to seven foundation models (Geneformer, scGPT, scCELLO, scPRINT, LangCell, UCE, scFoundation) on two datasets (Human Pancreas and Myeloid), achieving an average 98.7% reduction in attack success rates while preserving clean accuracy. OmicsDefense establishes a new paradigm for trustworthy single-cell analysis, enabling reliable cell atlas construction and clinical diagnosis in the era of foundation model-powered biomedicine.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 105