Noise Injection

Adds noise — random characters, extra spaces, punctuation, case randomization, Zalgo text — to disrupt keyword matching while preserving model readability. The model's training on noisy internet text provides robustness; keyword classifiers lack it.

• Input normalization (strip noise before classification)
• Subsequence extraction from noisy text

Chains with T2-AT-003 (Encoding) and T2-AT-004 (Unicode) for compound noise.