# Can I switch out a private RSA Key, so that when decrypting a message, it is decrypted to a different message?

Here's a task I've been struggling with: Say I have an encrypted message $$c$$, and a desired plaintext $$m$$. How can I define a valid private RSA Key $$(d,N)$$, so that $$c^d \bmod N = m$$

Note that the key needs to be valid! My approach was to write the modulo equation as $$y \times N + m = x \times N + c^d$$ so I get $$c^d-m'=(y-x) \times N$$. So I can simply set some value for $$d$$ and try to factorize the result to get $$N$$, but this factorization can be extremely hard because numbers are huge. Is there a better way to approach this?

• Be extremely precise with "valid". Is it required that $c<N$? If not, and $m\ll c$, what about for example $N=(c-m)/k$ for some $k\ge1$ and $d=1$ or $d=\lambda(N)+1$? Independently, is $N$ required to have no small factor, up to what? Is it required that $e$ in an $(e,N)$ matching $(d,N)$ be a small odd integer at least $3$? – fgrieu Apr 22 at 14:05
• valid means formally valid. E needs to match d, p and q need to be prime, (or probable prime). Practically, the key needs to pass OpenSSL RSA_check_key. – Kolja Vukolov Apr 22 at 14:06
• rsa_check_key tests are actually simple. I think c<N shouldnt matter for my task, but p and q need to be prime and p*q=N. – Kolja Vukolov Apr 22 at 17:44
• The tests in the OpenSSL documentation of RSA_check_key are indeed simple. But as for many things in OpenSSL, the documentation does not match the code (typically, the documentation does not change when the code does for either extending functionalities or fixing bugs). For example, the implementation allows 2048-bit $N$ with 3 prime factors, and up to 5 for larger $N$. And for two factors, it allows some keys with $d\,e\ne1\bmod((p-1)*(q-1))$ (because PKCS#1v2.2 allows it and FIPS 186-4 requires it for a large fraction of the keys). Check the source that I linked. – fgrieu Apr 22 at 17:57
• actually its supposed to work on the pycrypto implementation - I was more looking for a mathematical solution I guess – Kolja Vukolov Apr 22 at 18:30

Update: This other answer (temporarily hidden on request of the organizer of a CTF) details a different approach to the problem, working for all values of $$m$$ and $$c$$, yielding valid keys, that even are plausible except for a larger-than-customary $$e$$. These private keys will work nicely for those RSA implementations (including OpenSSL, pycrypto, and many others) accepting values of $$e$$ up to about $$N/2$$ (which is well within PKCS#1 limits). That's what the OP was asking for, rather than the following exploration.

In both crypto Catch-The-Flags and actual security attacks, it is important to determine what makes values accepted or not. Here two things matter most:

• Will something check that $$c (which is often a requirement in actual uses of RSA). A comment by the OP suggests there's no such check.
• What makes a private key unacceptable. Practice varies widely from most¹ to less stringent.

It happens that pycrypto (last indicated as applicable in a comment by the OP) is lenient on imported RSA keys, at least when using it's reference _slowcrypto.py code. See _importKeyDER and rsa_construct: as long as $$N$$, $$e$$, $$d$$, $$p$$ and $$q$$ are given, they are accepted as is. $$d_p$$, $$d_q$$, $$q_\text{inv}$$ are recomputed if imported from a key file or missing.

With such freedom, and $$c^d\bmod N=m$$ the only constraint involving $$c$$, if in the givens $$m (which I'll assume) then we can choose $$N=c-m$$ and $$d=1$$ (with other key parameters 42). The condition $$c^d\bmod N=m$$ gets met, success!

However, that can fail if the RSA private key is used in CRT form (using $$p$$, $$q$$, $$d_p$$, $$d_q$$, $$q_\text{inv}$$), which is common². We may be able to improve on this by searching a small $$k\in[1,c/m)$$ dividing $$c−m$$ and making $$N=(c−m)/k$$ odd and composite (if these are requirements). For $$p$$ and $$q$$, any decomposition as $$N=p\,q$$ with $$p>1$$, $$q>1$$, and $$\gcd(p,q)=1$$ should do. We'd set $$e=d=1$$, passing a test $$e\,d\equiv1$$ for any modulus. The key import will recompute $$d_p=d_q=1$$ and $$q_\text{inv}=q^{-1}\bmod p$$ as needed. The key should now work with or without using the CRT. I do not rule out that we could get away with $$p=1$$ or $$q=1$$ (with the other equal to $$N$$); or that, if we got blocked by a test that $$e\ge3$$, it could be circumvented with $$e=p\,q-p-q+2$$ or $$e=\lambda(N)+1$$.

¹ Testing that the key was generated per FIPS 186-4 would include verifying that $$N$$ has exactly 1024, 2048 or 3072 bits, $$N=p\,q$$ with $$p$$ and $$q$$ primes each in interval $$\left(2^{\left(\left\lceil\log_2(N)\right\rceil-1\right)/2},2^{\left\lceil\log_2(N)\right\rceil/2}\right)$$, $$e$$ is odd and in $$\left(2^{16},2^{256}\right)$$, $$d=e^{-1}\bmod\operatorname{lcm}(p-1,q-1)$$, $$d_p=e^{-1}\bmod(p-1)$$, $$d_q=e^{-1}\bmod(q-1)$$, $$q_\text{inv}=q^{-1}\bmod p$$. For 1024-bit $$N$$ the standard sets some extra requirement about $$p\pm1$$ and $$q\pm1$$, but these are never tested when accepting a key (contrary to the previous ones), because that would be very compute-intensive.

² We must however suppose that the advisable and common check that no fault crept in the computation of $$m$$ is skipped. That's performed by verifying $$c=m^e\bmod N$$, and that would fail even if we set an appropriate $$e$$ like $$e=1$$ or $$e=\lambda(N)+1$$. That would insure $$c\equiv m^e\pmod N$$, but not $$c=m^e\bmod N$$, since $$c\ge N$$ with our choice of $$N$$.