Remember that the dataset is often going to be small, and frequently will contain known or easily guessed strings. Uncompressed, often at predicable locations.

There are whole classes of cryptographic vulnerabilities which might result from either not compressing (compression should in theory would normalize entropy over the stream) or using a compression algorithm that results in a predictable dictionary, length, or other value in the same location (if junk padding isn't used properly).

Also, sending the state from client to server might open replay attacks and all sorts of other horrid situations.

Security depends on doing everything correctly all the time; this context IMO just feels open to too many plausible and unknown (potentially introduced in the future) vulnerabilities.