Pervasive Backdoor Vulnerabilities in Genomic Foundation Models
- 1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Abstract
Genomic foundation models are revolutionizing biological research and biotechnology, yet their vulnerability to data poisoning remains poorly characterized. Here we show that backdoor attacks can be reliably implanted during fine-tuning across diverse model architectures, tasks and model scales by poisoning only a small fraction of training data with short DNA motifs, including triggers derived from transposon terminal inverted repeats (TIRs). Attack success rates are generally high across encoder-based and autoregressive models, while clean-task performance remains almost unchanged. Increasing model scale does not confer measurable resistance. To detect such threats, we develop a two-stage auditing framework that flags potentially poisoned sequences through localized single-nucleotide mutation sensitivity in model predictions, without prior knowledge of trigger sequences. These results identify downstream fine-tuning as an attack surface for genomic foundation models and motivate stronger data provenance, adversarial evaluation and post-training security auditing.
Related
- Genomic compartmentalization of pervasive sex-biased gene expression in the vine mealybug Planococcus ficus
- Limitations of Genomic Foundation Models for Decoding Regulatory Mechanisms in ALS
- DNT: Diploid Genomic Foundation Model
- Genomic foundation models do not yet boost common variant discovery beyond standard functional annotations
- DAPE-PD: Diffusion adaptation of genomic foundation model enables tailored-generation of cis-regulatory DNA
BibTeX
@misc{ni2026pervasive,
title = {{Pervasive Backdoor Vulnerabilities in Genomic Foundation Models}},
author = {S. Ni and Q. Wang and C. Wei and X. Ni and S. Li and Z. Zhao and H. Li and R. Ji and R. Xu and H. Alinejad-Rokny and T. Wang and M. Yang},
howpublished = {bioRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.30.741642},
url = {https://www.biorxiv.org/content/10.64898/2026.07.30.741642v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - Pervasive Backdoor Vulnerabilities in Genomic Foundation Models
AU - S. Ni
AU - Q. Wang
AU - C. Wei
AU - X. Ni
AU - S. Li
AU - Z. Zhao
AU - H. Li
AU - R. Ji
AU - R. Xu
AU - H. Alinejad-Rokny
AU - T. Wang
AU - M. Yang
PY - 2026
DA - 2026/07/30
PB - bioRxiv
DO - 10.64898/2026.07.30.741642
UR - https://www.biorxiv.org/content/10.64898/2026.07.30.741642v2
AB - Genomic foundation models are revolutionizing biological research and biotechnology, yet their vulnerability to data poisoning remains poorly characterized. Here we show that backdoor attacks can be reliably implanted during fine-tuning across diverse model architectures, tasks and model scales by poisoning only a small fraction of training data with short DNA motifs, including triggers derived from transposon terminal inverted repeats (TIRs). Attack success rates are generally high across encoder-based and autoregressive models, while clean-task performance remains almost unchanged. Increasing model scale does not confer measurable resistance. To detect such threats, we develop a two-stage auditing framework that flags potentially poisoned sequences through localized single-nucleotide mutation sensitivity in model predictions, without prior knowledge of trigger sequences. These results identify downstream fine-tuning as an attack surface for genomic foundation models and motivate stronger data provenance, adversarial evaluation and post-training security auditing.
ER -