Predicting Privacy Leakage from Weight Spectral Density
Cheap spectral signals may flag models that leak membership data.
The paper tests whether WeightWatcher metrics can stand in for costly shadow-model membership inference audits. On image and tabular classification tasks, stable rank tracks positively with overall attack success. Log alpha-Norm is consistently negative against vulnerability in the low false-positive regime. The authors say these links beat the generalisation gap, suggesting model spectra capture privacy risk that overfitting metrics miss. ArXiv · AI/CL/LG's note
The paper tests whether WeightWatcher metrics can stand in for costly shadow-model membership inference audits. On image and tabular classification tasks, stable rank tracks positively with overall attack success. Log alpha-Norm is consistently negative against vulnerability in the low false-positive regime. The authors say these links beat the generalisation gap, suggesting model spectra capture privacy risk that overfitting metrics miss. ArXiv · AI/CL/LG's note
score 4