what does it mean the distinguisher in LWE problem in which the secret key changes constantly for each ciphertext and plaintext pair?
this isn't true, so if you have a particular source that is claiming that and leading to your confusion you should probably link it. The correct definitions are below.
The LWE problem can be stated as follows.
LWE Distribution: For $\vec s\in\mathbb{Z}_q^n$, let $\mathsf{LWE}_{n,q,\chi}(\vec s)$ be the distribution that samples $\vec a\gets \mathbb{Z}_q^n$ and $e\gets \chi$, and then outputs $(\vec a, \langle \vec a,\vec s\rangle + e)$.
Here $\chi$ is a "concentrated" distribution on $\mathbb{Z}$.
Typical examples are discrete Gaussians, or bounded uniform cdistributions $\{-B,-B+1,\dots, B-1,B\}$, though there are many choices possible.
LWE Problem: An adversary is given query access to an oracle $\mathcal{O}$ that is either
- Uniform: Uniform samples $(\vec a, u)\gets \mathbb{Z}_q\times \mathbb{Z}$, or
- LWE: For $\vec s\gets \mathbb{Z}_q^n$, the oracle $\mathsf{LWE}_{n,q,\chi}(\vec s)$,
And must determine which oracle they are given.
This is to say that the secret $\vec s$ is fixed. Note that sampling $\vec s\gets \mathbb{Z}_q^n$ is one choice, but one can also sample $\vec s\gets \chi^n$ without degrading the security of the problem (and in applications to homomorphic encryption this is preferrable).
The fact that $\vec s$ is fixed is vital to be able to prove security of cryptosystems. Without this assumption, you could prove security to encrypt a single integer, but I don't see how you could do much else, e.g. the cryptosystem would be "one-shot".