It is said that the goal of cryptography is to make the attacker resort to rubber-hose cryptanalysis, revealing their intentions. In those cases, the best you can hope for is plausible deniability, i.e. decrypt some part of the cipher showing something embarrassing but not illegal, and hope that they think that was all that was there. Actually that's what TrueCrypt provides with its "secret encrypted volume". However, there are some tests for randomness (http://en.wikipedia.org/wiki/Randomness_tests) like the chi-squared test which can detect/identify many such encryption schemes if the size is large enough.
In short ... it's an arms race. Your best bet is Steganography (http://en.wikipedia.org/wiki/Steganography) which has been around for centuries. Just write messages where no one suspects. You can encrypt them with PFS and plausible deniability for best effect.
For more information, see "Is it possible to distinguish a securely-encrypted ciphertext from random noise?" at http://crypto.stackexchange.com/q/1646/2454
That would be an insane scenario.
(I am not a serious cryptographer)
There is literally no way to keep a conversation secret if one of the parties involved is compromised, because you're talking to a compromised party!
A snooping agency holding the private key of, say, Google would still need to intercept each connection and act as a man in the middle. Simply knowing the RSA key (either beforehand or afterwards) is not enough to decrypt PFS sessions.
*edit: Google would still be able to collect and hand over all session keys ever used to communicate with them. The NSA might then use those session keys to decrypt any corresponding session.