Privacy Adversarial Network: Representation Learning for Mobile Data Privacy
summary
Introduces PAN, a device-side encoder trained with a utility discriminator and one or two adversaries to force representations to keep task-relevant signals while hiding sensitive ones. Evaluated on six datasets across digit recognition, image, audio, activity and driver behavior tasks, PAN (PAN_1 and PAN_2 with two attackers) outperforms four baselines, yields higher task accuracy than a standard DNN, and increases reconstruction error for privacy, claiming an automatic encoder deployable on-device.
AI-written summary · about
related talks
43mPanel: Privacy Design Patterns for AI Systems: Threats and Protections
14mEstablishing Privacy Metrics for Genomic Data Analysis
05mPMC: A Privacy-preserving Deep Learning Model Customization Framework for Edge Computing
31mIs your model private?
38mRepresentation via Representations: Improving Generalization to Unseen Populations
40mNew Face, Who Dis? Protecting Privacy in a World of Surveillance
44mPrivacy by Design Machine Learning
28mPrivacy preserving machine learning