Hardware Wallet attacks by input ownership omission and fix

----------------------------------------------------------------------------------

So a while back I realized that to have HW wallets do safe automated coinjoins(without any user interaction be sure there are no fee dumps or handing money to others) you have to protect yourself from the case of signing one set of inputs while the other owned set is hidden from the device, then repeating the same action with the two sets reversed.

Note that there is no support for such a mode in HW wallets today, but could possibly greatly increase liquidity of JoinMarket like systems.

First signing pass:

1 BTC (yours, host tells ledger about it) --------

                                                >   1.5 BTC

1 BTC (yours, host fails to tell ledger about it)-

Second signing pass:

1 BTC (yours, host fails to tell ledger) ---------

                                                >   1.5 BTC

1 BTC (yours, host tells ledger about it)---------

In this scenario, you sign the first input, thinking "great I'm getting 0.5 BTC for running coinjoin" when in reality this will simply be re-played again later with the inputs switched, *costing* you 0.5 BTC. (Ledger doesn't support "negative fees", but imagine more more inputs are included that aren't yours.)

More recently I noticed a more common issue along the same lines:

With Segwit inputs, the entire transaction referred to in the prevout is generally no longer included for HW wallet signing API. This greatly speeds up signing since potentially multiple MBs of transactions are no longer passed into the device, but comes with a cost: An attacker can claim certain inputs' value is much lower than it actually is. In the first pass, the host reports the first input's value properly, and the second as lower. The signature on the first input will go through fine(value included in the sighash is only for that input), then attacker prompts a restart of signing, reporting the 2nd value properly, and first value improperly low, which allows the attacker to report the same fee twice on the device. Both signatures over each input are correct, but the user was prompted with an invalid fee amount(too low).

To fix this I consulted with andytoshi and got something we think works for both cases:

1) When a signing device receives a partially signed transaction, all inputs must come with a ownership proof:

- For the input at address A, a signature over H(A || x) using the key for A. 'x' is some private fixed key that only the signing device knows(most likely some privkey along some unique bip32 path).

- For each input ownership proof, the HW wallet validates each signature over the hashed message, then attempts to "decode" the hash by applying its own 'x'. If the hash doesn't match, it cannot be its own input.

- Sign for every input that is yours

This at a minimum makes sure that the wallet's total "balance" will not go down more than the reported fee.

Benefits:

- Still small memory footprint compared to legacy signing

- Allows user-interactionless coinjoins without putting funds at risk

- These proofs can be created at any time, collected at the front of any CoinJoin like protocol.

- These proofs can be passed around as additional fields for Partially Signed Bitcoin Transactions: https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-August/014838.html