Representation Matters in Randomized Smoothing for Audio Classification
Researchers have identified a critical ambiguity in applying randomized smoothing for audio classification robustness certification. The standard method assumes noise is added in a single vector space, but audio processing often involves multiple transformations, making the certified object unclear. The study demonstrates that different representations, such as raw waveforms versus log-mel features, and preprocessing steps like normalization, significantly alter the robustness certification results. The authors recommend explicit reporting of the certified object, perturbation model, and post-noise geometry changes for accurate and reproducible audio robustness studies. AI
IMPACT Clarifies methodology for certifying AI model robustness in audio tasks, crucial for safety-critical applications.