Connectome-based machine learning models are vulnerable to subtle data manipulations.

Area of Science:

• Neuroimaging
• Machine Learning
• Data Science

Background:

• Neuroimaging-based predictive models are advancing.
• Model trustworthiness (robustness to data manipulation) is often overlooked.
• High trustworthiness is crucial for reliable research findings.

Purpose of the Study:

• To investigate the impact of minor data manipulations on machine learning predictions using functional connectomes.
• To assess methods for falsely enhancing prediction performance.
• To evaluate adversarial noise attacks designed to degrade performance.

Main Methods:

• Utilized functional connectomes for analysis.
• Introduced data manipulations including performance enhancement methods and adversarial noise attacks.
• Compared original and manipulated data using similarity metrics (r = 0.99).

Main Results:

• Minor data manipulations significantly altered machine learning prediction performance.
• Manipulated data remained highly similar to original data (r = 0.99).
• These manipulations did not impact other downstream analyses.

Conclusions:

• Functional connectome data can be subtly modified to achieve desired prediction outcomes.
• Existing adversarial attacks and novel enhancement attacks pose risks to model trustworthiness.
• Developing countermeasures is essential to maintain the integrity of neuroimaging research and its applications.