1. Proceeding Structure

2. The Stem of the Paper

3. Points need attention

4. Proposed Attack

5. Experimental Setting

6. Experiments

Conclusion

Following direction.

1. Attacking aspect. Attacking means may improve from some aspects, e.g. GAN optimization, or other information extraction through generative methods.
2. The local and the global. In federated learning, there are differences between local models and the global model. So the optimization of some methods focused on the relationship between local parameters and global parameters may be feasible.
3. Evaluation standard(cv, etc). The concrete standard may adapt to various fields, e.g. CV, NLP, etc. A judge mechanism, which has two levels. Firstly, whether there is a privacy offense. Secondly, whether the attack is effective.
4. Theoretical proof also needs great attention. What effect(denote with probability) can we achieve under what constraint.