Safeguarding open-weight genomic foundation models through weight locking

Area of Science:

• Genomic foundation models
• Artificial Intelligence in Biology
• Biosecurity

Background:

• Genomic foundation models accelerate biological research by learning transferable representations.
• Excluding viral genomes from training data is a common but circumventable biosecurity measure.
• Weight-locking with spectral deformation is a proposed defense against model fine-tuning.

Purpose of the Study:

• To evaluate the effectiveness of spectral deformation locking against various fine-tuning attacks on a genomic foundation model.
• To assess the model's vulnerability to both naive and informed adversarial attacks.

Main Methods:

• Applied spectral deformation locking to the Evo-1-8k-base genomic foundation model.
• Tested naive fine-tuning, low-rank adaptation (LoRA), and singular value decomposition (SVD)-chain attacks.
• Quantified recovered virological capabilities on Human Virome Understanding Evaluation (HVUE) tasks.

Main Results:

• The lock successfully defended against naive fine-tuning and LoRA attacks.
• Naive attacks resulted in capability loss, not gain, on pathogenicity and host tropism.
• Informed SVD-chain attacks recovered pathogenicity prediction capabilities at increased computational cost.

Conclusions:

• Spectral deformation locking is effective against naive fine-tuning attacks on genomic foundation models.
• Advanced adversarial attacks can still compromise model security, highlighting the need for robust defenses.
• The findings inform strategies for developing secure and reliable AI models in biological research.