English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

We analytically investigate how over-parameterization of models in randomizedmachine learning algorithms impacts the information leakage about theirtraining data. Specifically, we prove a privacy bound for the KL divergencebetween model distributions on worst-case neighboring datasets, and explore itsdependence on the initialization, width, and depth of fully connected neuralnetworks. We find that this KL privacy bound is largely determined by theexpected squared gradient norm relative to model parameters during training.Notably, for the special setting of linearized network, our analysis indicatesthat the squared gradient norm (and therefore the escalation of privacy loss)is tied directly to the per-layer variance of the initialization distribution.By using this analysis, we demonstrate that privacy bound improves withincreasing depth under certain initializations (LeCun and Xavier), whiledegrades with increasing depth under other initializations (He and NTK). Ourwork reveals a complex interplay between privacy and depth that depends on thechosen initialization distribution. We further prove excess empirical riskbounds under a fixed KL privacy budget, and show that the interplay betweenprivacy utility trade-off and depth is similarly affected by theinitialization.

Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networks