Major Description
Gary Gregory
Aug 5, 2020, 10:53 PM (20 hours ago)
to Commons
Hi All:
I am seeing what may be a random AEADBadTagException in GcmCipherTest?
For example:
[ERROR] testGcmTamperedData(org.apache.commons.crypto.cipher.GcmCipherTest)
Time elapsed: 0.015 s <<< ERROR!
881java.lang.Exception: Unexpected exception,
expected<javax.crypto.AEADBadTagException> but
was<java.lang.InternalError>
882 at org.apache.commons.crypto.cipher.GcmCipherTest.testGcmTamperedData(GcmCipherTest.java:224)
883
884
Any thoughts?
The above is from
https://travis-ci.org/github/apache/commons-crypto/jobs/715348986
Gary
Alex Remily
8:10 AM (11 hours ago)
to Commons
That is an intermittent issue that I haven't been able to reliably reproduce. As I recall, the test that's failing is supposed to fail, but in a different way. I think it's supposed to fail because of a short buffer but occasionally fails because of an internal error, and when that happens this test fails. I don't know when it was introduced. We should probably document it in jira and or realese notes.
Matt Sicker
10:33 AM (8 hours ago)
to Commons
Now I hope we don't have unit tests depending on non-static state for
its random number generator! 
I'd expect a crypto library's test
suites to include several hard-coded known-good and known-bad
ciphertexts with static keys/IVs similar to the test cases presented
in their RFCs (especially since said tests are typically small enough
to copy/paste the binary data fairly easily).
–
Matt Sicker
Rob Tompkins
10:37 AM (8 hours ago)
to Commons
We actually do have a considerable number of those in our projects where we use probabilistic epsilons on the output. See commons-rng. Note, Gilles is quite good at writing such tests.
-Rob
Matt Sicker
10:42 AM (8 hours ago)
to Commons
Well, for testing RNGs, I can understand using property testing, yes.
It would also be useful for testing fuzzing scenarios like making sure
the GCM tag is invalid for any random input data (giving a near zero
probability of valid data) or that an elliptic curve implementation
doesn't leak out information about points outside the curve or respond
to invalid inputs improperly or things like that.
Rob Tompkins
10:50 AM (8 hours ago)
to Commons
+1 - the elliptic curve stuff I’ll have to defer to you on as I’m less a number theorist and more of a logician.
-Rob
Gary Gregory
10:56 AM (8 hours ago)
to Commons
This is all fine and good but how would you fix the test such that it does
not fail randomly. PR anyone?
Gary
Rob Tompkins
10:57 AM (8 hours ago)
to Commons
Either static inputs for determinism, or putting a probabilistic boundary in which the solution can fall.
-Rob
Matt Sicker
11:00 AM (8 hours ago)
to Commons
Choose a seed value for the `new Random()` constructor and the tests
will be deterministic.
Matt Sicker
11:02 AM (8 hours ago)
to Commons
Or alternatively, if using random values each time, have it retry the
test with a different value. It's typically better to use an actual
property testing library for these types of tests anyways. One example
library I found is https://jqwik.net/ (these types of testing
libraries are more common in functional programming like in Scala).
Rob Tompkins
11:03 AM (8 hours ago)
to Commons
Precisely. That’s another technique we’ve used in rng.
-Ropb
