Membership Inference Attacks against Machine Learning Models

[@tjr3117]

Thank you for sharing this video. It's good to study about privacy leakage in machine learning with sensitive data.

[@valentin8482]

Outstandingly good explanation of the topic. Thanks a lot

[@top_coder]

Great talk! I learned a a lot.

[@simoncarter3541]

What is the difference between the train and test data for the shadow models? As they're labelled "in" and "out" respectively?

[@silentrandom]

Very well explained. Is there any talk related to how privacy-preserving ml e.g. Homomorphic encryption, mpc technique countermeasure such attacks?

[@tjr3117]

In my opinion, i'm now studying privacy preserving methods, the best thing against this attack is differential privacy. PHE is too slow, so many research to faster encryption performance with gpu parallel operation is going on.

[@sarbajit.g]

This attack is implemented over too many assumptions. If somebody knows a similar model architecture and can collect similar types of datasets why he might bother to attack a model. Also to launch an MIA, to only know, that a particular person's data, is used during training in an ML model seems like absolute overwork. Is there any real-world example of MIA till data or it is just for an academic purpose?

[@s4098429]

Terrible audio, very hard to understand.

[@高天林-r2w]

Good paper, Great talk

[@Orqngesyck]

Great Talk

[@ramizkarim9037]

CANT UNDERSTAND THE ACCENT

[@chaoliu4328]

very clear

[@hamidfazli6936]

Very clever!