I found a counterexample to this problem.
Suppose that $F':\{0,1\}^n\times\{0,1\}^n\rightarrow \{0,1\}^n$ is SPRP.
Then define $F:\{0,1\}^n\times\{0,1\}^n\rightarrow \{0,1\}^n$ as follows: Let $F_{0^n}():=F'_{0^n}()$. For $k\not=0^n$, we define:
$$
F_k(x):=\begin{cases}
F'_k(0^n)\oplus k &\text{ if }x=k,\\
F'_k(k) &\text{ if }x=F'^{-1}_k(F'_k(0^n)\oplus k),\\
F'_k(x) &\text{otherwise}.
\end{cases}
$$
Actually I just interchange the values of $F_k'$ at $k$ and $F'^{-1}_k(F'_k(0^n)\oplus k)$.
Now if we define $\hat{F}_k(x):=F_k(k)\oplus F_k(x)$, we can see $\hat{F}_k(0^n)=k$ for any $k\not=0^n$, so clearly $\hat{F}$ is not a PRP.
It remains to prove that $F$ constructed above is a SPRP. Suppose that there is an efficient distinguisher $A$ that can break the SPRP property of $F$, we construct a $D$ to break the SPRP property of $F'$:
$D$ is given access to oracles $O()$ and $O^{-1}()$, where $O()=F'_k()$ or a uniform random permutation$f()$.
- $D$ answer $A$'s querys at $x_1,x_2,...,x_q$ using $O()$ and $O^{-1}()$. After $q$ times query $A$ output a bit $b$.
- Then $D$ output $b$.
If $D$ is given access to $f,f^{-1}$, then $D$ answers $A$'s queries correctly. So $Pr[D^{f,f^{-1}}(1^n)=1]=Pr[A^{f,f^{-1}}(1^n)=1]$.
If $D$ is given access to $F'_k, F'^{-1}_k$, let event $E:= A \text{ queries at } k,F'^{-1}_k(F'_k(0^n)\oplus k),F'_k(0^n)\oplus k,F'_k(k)$.
Then $D$ answers $A$'s
queries correctly if $E$ does not occur! If $Pr(E)$ is negl, then the difference between $Pr[D^{F'_k, F'^{-1}_k}(1^n)=1]$ and $Pr[A^{F_k, F^{-1}_k}(1^n)=1]$ is negl, then the advantage of $D$ is non-negl.
If $Pr(E)$ is non-negl, intuitively, we can use these queries to find the key with non-negl probability.
Formally we construct $D'$:
- $D'$ answer $A$'s querys at $x_1,x_2,...,x_q$ using $O()$ and $O^{-1}()$.
- $D'$ randomly choose $x_i$ from $A$'s $q$ queries.
- $D'$ computes $k_1=x_i, k_2=O(x_i)\oplus O(0^n),k_3=x_i\oplus O(0^n)$,$k_4=O^{-1}(x_i)$.
- Then $D'$ output 1 if $O()=F'_{k_1} $ or $F'_{k_2}$ or $F'_{k_3}$ or $F'_{k_4}$. ($D'$ can check if $O()=F'_k$ for any $k$ with high accuracy). Otherwise $D'$ output $0$.
It is clear that $Pr[D'^{f,f^{-1}}(1^n)=1]$ is negl. On the other side, $Pr[D'^{F'_k, F'^{-1}_k}(1^n)=1]\ge 1/q*Pr(E)$ is non-negl.