Please enable JavaScript.
Coggle requires JavaScript to display documents.
Membership Inference Attacks, Adversarial Attack against ML, Federated…
Membership Inference Attacks
Attack Goal
Against User's privacy
Use in a good way
Data auditing
Has User-A's data been used by the cloud provider to train a ML model
Unlearning & earssing
Adversarial Attack against ML
In image case
High-confidence region in Feature space
Salienty map
Attack's consistent miscalssification to detect the Adv. aatack
Teacher-Student Model
Mixup Traning
Different function like non-conex
input permutation
Federated Learning
Secure Aggragation
Homeoorphic is better than Multi-party computation
Prototype-based Federated learning
Prototype is a sum of the representation learned by a model
Subject to source-membership Inference