Universal and transferable attacks on pathology foundation models using microscopic perturbations.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Pathology
• Digital Microscopy

Background:

• Foundation models have revolutionized pathology and optical microscopy.
• These advanced AI systems are vulnerable to adversarial attacks, posing potential risks.

Purpose of the Study:

• To introduce Universal and Transferable Adversarial Perturbations (UTAP) for pathology foundation models.
• To reveal critical vulnerabilities and assess the impact of adversarial attacks on AI in microscopy.

Main Methods:

• Developed UTAP using deep learning, creating a fixed, weak microscopic noise pattern.
• Applied UTAP to pathology images, disrupting feature representations of foundation models.
• Evaluated UTAP across diverse datasets and state-of-the-art pathology foundation models.

Main Results:

• UTAP caused significant performance drops in downstream tasks, including misclassification across unseen data.
• Demonstrated universality: UTAP is effective across different datasets and fields-of-view.
• Showcased transferability: UTAP degraded performance of unseen, external black-box models.

Conclusions:

• UTAP represents a broad threat to pathology foundation models, not specific to any single model or dataset.
• These findings establish a benchmark for AI model robustness in pathology.
• Emphasized the need for advanced defense mechanisms for safe AI deployment in microscopy and pathology.