Jalpesh Vasa ,

The answer is at most points that are possible. You must look at an attacker as an individual that will follow the path of least resistance, that is, the weakest link in the whole process. You must also consider that an attacker has time on his side and his attack methodology is bound by the resources at his disposal.

Given this, what you can do is analyse the particular implementation and develop the security with respect to what is the likeliest path of attack given the resources of the attacker and make it hard enough such that the work function of the attacker exceeds the value of the asset. Privacy analysis is just one subcategory on this process of analysis.

There is no single defence strategy that will cover all possible scenarios. You must design according to particular implementations.

There are many possible privacy threats to the deep learning model span over the model input data, hyperparameters, communication, and inference queries from the model output.

I recommend you read the following survey on privacy in deep learning https://arxiv.org/pdf/2004.12254.pdf

In the European Union privacy is on stake when the data subject has not given consent for the purpose you use his personal data. This has to be agreed outside your deep learning application. If not you violate European Data Protection law applying what ever data analysis on personal data without explicit consent of the data subject.

It must be applied on the top level of the data source application