Random Erasing vs. Model Inversion: A Promising Defense or a False Hope?
Researchers have discovered that Random Erasing (RE), a technique typically used to improve model generalization, can also serve as an effective defense against model inversion attacks. These attacks aim to reconstruct private training data from machine learning models, posing a significant privacy risk. The study found that RE introduces a discrepancy in the feature space, degrading the quality of reconstructed data and reducing attack accuracy while maintaining model utility. The effectiveness of RE is influenced by factors like partial erasure and random location of the erased regions. AI
IMPACT Introduces a simple, effective defense against data privacy attacks in ML models, potentially improving trust and adoption.