Let $m \in \{0, 1\}^{2n} = m_1 || m_2$. Let's also assume that $F_k$ is a PRF and $G$ is a PRG defined as $G: \{0, 1\}^n \rightarrow \{0, 1\}^n$. Now, Let's define the $MAC$ scheme as follows:
$$MAC(m) = F_{m_1}(k \oplus G(k\oplus m_2))$$
Where $k \in \{0, 1\}^n$. I think it's correct, but I have not seen much examples of using the message itself as PRF's key. That is why I am getting confused. Is it correct and/or secure?
No. The construction is not secure!
The security property of MAC says that, every poly-time adversary who has access to a MAC oracle (the oracle outputs MAC on input message) cannot come up with a MAC on a message of his choice (without querying the oracle) with non-negligible probability. In your construction, suppose you know that $G(x) = G(x')$ for some known values $x, x'$. Then an adversary can query for MAC on $m_1 || x$ (for any message $m_1$) and then obtain MAC for $m_1 || x'$.
Note: It's easy to construct a PRG $G$ such that $G(x) = G(x')$. Take any PRG $G'$ and pick any two values $x,x'$. Construct $G$ as follows. $G(y) = G'(y)$ if $y \neq x$. $G(y) = G'(x')$ if $y = x$. Note that $G$ is a PRG if $G'$ is a PRG.
