SPV clients don't fully validate blocks. Therefore, they are vulnerable to being fed a fake block from their full node peers. A fraud proof is a way to prove that a chain tip is invalid. As soon as any node would discover a block to be invalid, they'd create and relay a fraud proof to the network. As SPV clients receive the fraud proof they will know not to accept the invalid block.
As hinted above, fraud proofs can be created by any network participant that discovers an invalid block. As far as I know, this has nothing to do with Segregated Witness, but will rather be introduced as an independent softfork some time in the future.
what incentivises a full node to publish said fraud proof, other than it being out of genuine kindness? And what happens if a SPV node encounters conflicting fraud proofs where 2 partitions in the network sends differing fraud proofs that invalidates either chain tips? – renlord – 2016-11-15T13:17:41.203
If there are two competing chain tips, they are both valid. Fraud proofs are only used to identify invalid blocks. Reducing possibilities for users to be scammed on the network increases the value of the network for all users and there is practically zero effort involved for the generator of the fraud proof. I.e. it would still make sense for a completely self-interested user to publish a fraud proof. – Murch – 2016-11-15T13:23:08.937
Now there is a BIP for oversize blocks: https://github.com/bitcoin/bips/blob/master/bip-0180.mediawiki
– Janus Troelsen – 2017-05-29T11:08:55.483