Gradient-based enhancement attacks in biomedical machine learning.

Area of Science:

• Biomedical research
• Machine learning
• Data science

Background:

• Machine learning is increasingly used in biomedical research, but its trustworthiness is often unexamined.
• Adversarial attacks can degrade model performance, but enhancement attacks pose a greater threat by falsely inflating performance.

Approach:

• Developed two techniques for enhancing classifier prediction performance with minimal feature changes.
• Demonstrated general performance enhancement and method-specific performance improvement.

Key Points:

• Enhancement framework falsely boosted accuracy from 50% to nearly 100% with high feature similarity (Pearson's r > 0.99).
• Method-specific enhancement falsely favored one model over another (e.g., neural network over logistic regression by 17%).
• Original and enhanced data remained highly similar (r = 0.99).

Conclusions:

• Minor data manipulations can achieve any desired prediction performance, posing ethical challenges for biomedical machine learning.
• Emphasizes the need for robust data provenance tracking and precautionary measures to ensure research integrity.