Noise Injection
Deliberate introduction of random perturbations into training data or model parameters to enhance robustness and guard against adversarial manipulation.
Definition
A regularization strategy where Gaussian noise is added to inputs, hidden activations, or weights during training. It helps models generalize by preventing over-reliance on specific patterns and increases resilience to input perturbations. Governance requires monitoring the impact on performance, ensuring injected noise levels align with deployment scenarios, and documenting noise parameters for audit and retraining consistency.
Real-World Example
To harden its speech-recognition AI against background disturbances, a voice-assistant developer injects varying levels of white noise into training audio. Post-training tests show word-error rates degrade by only 2% under real-world noise conditions—compared to 10% without noise injection—demonstrating improved robustness.
.svg)
