Timezone: »

DP-InstaHide: Data Augmentations Provably Enhance Guarantees Against Dataset Manipulations
Eitan Borgnia · Jonas Geiping · Valeriia Cherepanova · Liam Fowl · Arjun Gupta · Amin Ghiasi · Furong Huang · Micah Goldblum · Tom Goldstein
Data poisoning and backdoor attacks manipulate training data to induce security breaches in a victim model. These attacks can be provably deflected using differentially private (DP) training methods, although this comes with a sharp decrease in model performance. The InstaHide method has recently been proposed as an alternative to DP training that leverages supposed privacy properties of the mixup augmentation, although without rigorous guarantees. In this paper, we rigorously show that $k$-way mixup provably yields at least $k$ times stronger DP guarantees than a naive DP mechanism, and we observe that this enhanced privacy guarantee is a strong foundation for building defenses against poisoning.

Author Information

Eitan Borgnia (University of Maryland)
Jonas Geiping (University of Maryland, College Park)
Valeriia Cherepanova (University of Maryland)
Liam Fowl (University of Maryland)
Arjun Gupta (University of Maryland, College Park)
Amin Ghiasi (University of Maryland, College Park)
Furong Huang (University of Maryland)
Micah Goldblum (New York University)
Tom Goldstein (University of Maryland)

More from the Same Authors