Our main contributions are: 1.

Flexible variational distributions improve variational inference but are harder to optimize.

However,untrusted data sources leavethe services vulnerable to poisoning attacks [5, 28], where adversaries can inject malicious training data to degrade model accuracy.