Stress Testing

Stress Testing for ML models is the systematic evaluation of model performance under extreme or challenging conditions — pushing inputs beyond typical operating ranges to identify failure modes, performance degradation, and the limits of reliable model operation.

Stress Testing Approaches

• Distribution Shift: Test on data from different distributions (different fab, different product, different time period).
• Extreme Values: Feed inputs at the boundaries or beyond the training data range.
• Noise Injection: Add increasing levels of noise to inputs to find the noise threshold for failure.
• Adversarial: Apply adversarial perturbations of increasing strength.

Why It Matters

• Failure Discovery: Stress testing reveals failure modes invisible in standard accuracy evaluation.
• Operating Envelope: Defines the reliable operating envelope of the model — where it can and cannot be trusted.
• Production Safety: Models deployed in semiconductor fabs must be tested under stress before controlling real processes.

Stress Testing is pushing the model to its limits — finding where and how the model breaks to ensure safe deployment.