Some comments: some purely editorial, some substantive.
Editorial: stuff is xored with zero, the concatenation language is not used consistently. I found it difficult to understand the proposed scheme and check equivalence to the paper. Maybe some more words to explain the layering would help.

Substantive: Does it matter that it is possible to compute a message that doesnt change the digest if you know the key?

On Fri, Mar 1, 2019 at 9:05 AM Nick Mathewson <nickm@torproject.org> wrote:
>
> Hi!
>
> I'm sending a new version of proposal 295 from Tomer Ashur, Orr
> Dunkelman, and Atul Luykx.  It's an updated version of their design
> for an improved relay cell encryption scheme, to prevent tagging
> attacks.
>
> This proposal is checked into the torspec repository.  I'm also
> linking to a diagram for this scheme (and its latex source) from Atul
> Luykx: https://people.torproject.org/~nickm/prop295/
>
> Finally, I have a draft python reference implementation for an older
> version of this proposal.  I hope to be updating it soon and sending
> out a link next week.
>
> cheers!  -- Nick
>
>
>
> Filename: 295-relay-crypto-with-adl.txt
> Title: Using ADL for relay cryptography (solving the crypto-tagging attack)
> Author: Tomer Ashur, Orr Dunkelman, Atul Luykx
> Created: 22 Feb 2018
> Last-Modified: 1 March 2019
> Status: Open
>
>
> 0. Context
>
>    Although Crypto Tagging Attacks were identified already in the
>    original Tor design, it was not before the rise of the
>    Procyonidae in 2012 that their severity was fully realized. In
>    Proposal 202 (Two improved relay encryption protocols for Tor
>    cells) Nick Mathewson discussed two approaches to stymie tagging
>    attacks and generally improve Tor's cryptography. In Proposal 261
>    (AEZ for relay cryptography) Mathewson puts forward a concrete
>    approach which uses the tweakable wide-block cipher AEZ.
>
>    This proposal suggests an alternative approach to Proposal 261
>    using the notion of Release (of) Unverified Plaintext (RUP)
>    security. It describes an improved algorithm for circuit
>    encryption based on CTR-mode which is already used in Tor, and an
>    additional component for hashing.
>
>    Incidentally, and similar to Proposal 261, this proposal employs
>    the ENCODE-then-ENCIPHER approach thus it improves Tor's E2E
>    integrity by using (sufficient) redundancy.
>
>    For more information about the scheme and a security proof for
>    its RUP-security see
>
>        Tomer Ashur, Orr Dunkelman, Atul Luykx: Boosting
>        Authenticated Encryption Robustness with Minimal
>        Modifications. CRYPTO (3) 2017: 3-33
>
>    available online at https://eprint.iacr.org/2017/239 .
>
>    For authentication between the OP and the edge node we use
>    the PIV scheme: https://eprint.iacr.org/2013/835
>
> 2. Preliminaries
>
> 2.1 Motivation
>
>    For motivation, see proposal 202.
>
> 2.2. Notation
>
>    Symbol               Meaning
>    ------               -------
>    M                    Plaintext
>    C_I                  Ciphertext
>    CTR                  Counter Mode
>    N_I                  A de/encryption nonce (to be used in CTR-mode)
>    T_I                  A tweak (to be used to de/encrypt the nonce)
>    T'_I                 A running digest
>    ^                    XOR
>    ||                   Concatenation
>           (This is more readable than a single | but must be adapted
>           before integrating the proposal into tor-spec.txt)
>
> 2.3. Security parameters
>
>    HASH_LEN -- The length of the hash function's output, in bytes.
>
>    PAYLOAD_LEN -- The longest allowable cell payload, in bytes. (509)
>
>    DIG_KEY_LEN -- The key length used to digest messages (e.g.,
>    using GHASH). Since GHASH is only defined for 128-bit keys, we
>    recommend DIG_KEY_LEN = 128.
>
>    ENC_KEY_LEN -- The key length used for encryption (e.g., AES). We
>    recommend ENC_KEY_LEN = 128.
>
> 2.4. Key derivation (replaces Section 5.2.2)
>
>    For newer KDF needs, Tor uses the key derivation function HKDF
>    from RFC5869, instantiated with SHA256. The generated key
>    material is:
>
>                  K = K_1 | K_2 | K_3 | ...
>
>    where, if H(x,t) denotes HMAC_SHA256 with value x and key t,
>          and m_expand denotes an arbitrarily chosen value,
>          and INT8(i) is an octet with the value "i", then
>              K_1     = H(m_expand | INT8(1) , KEY_SEED )
>          and K_(i+1) = H(K_i | m_expand | INT8(i+1) , KEY_SEED ),
>    in RFC5869's vocabulary, this is HKDF-SHA256 with info ==
>    m_expand, salt == t_key, and IKM == secret_input.
>
>    When used in the ntor handshake a string of key material is
>    generated and is used in the following way:
>
>    Length       Purpose                         Notation
>    ------        -------                        --------
>    HASH_LEN     forward digest IV               DF      *
>    HASH_LEN     backward digest IV              DB      *
>    ENC_KEY_LEN  encryption key                  Kf
>    ENC_KEY_LEN  decryption key                  Kb
>    DIG_KEY_LEN  forward digest key              Khf
>    DIG_KEY_LEN  backward digest key             Khb
>    ENC_KEY_LEN  forward tweak key               Ktf
>    ENC_KEY_LEN  backward tweak key              Ktb
>    DIGEST_LEN   nonce to use in the                      *
>                   hidden service protocol
>
>       * I am not sure that we need these any longer.
>
>    Excess bytes from K are discarded.
>
> 2.6. Ciphers
>
>    For hashing(*) we use GHASH with a DIG_KEY_LEN-bit key. We write
>    this as Digest(K,M) where K is the key and M the message to be
>    hashed.
>
>    We use AES with an ENC_KEY_LEN-bit key. For AES encryption
>    (resp., decryption) we write E(K,X) (resp., D(K,X)) where K is an
>    ENC_KEY_LEN-bit key and X the block to be encrypted (resp.,
>    decrypted).
>
>    For a stream cipher, unless otherwise specified, we use
>    ENC_KEY_LEN-bit AES in counter mode, with a nonce that is
>    generated as explained below. We write this as Encrypt(K,N,X)
>    (resp., Decrypt(K,N,X)) where K is the key, N the nonce, and X
>    the message to be encrypted (resp., decrypted).
>
>    (*) The terms hash and digest are used interchangeably.
>
> 3. Routing relay cells
>
> 3.1. Forward Direction
>
>    The forward direction is the direction that CREATE/CREATE2 cells
>    are sent.
>
> 3.1.1. Routing from the Origin
>
>    Let n denote the integer representing the destination node. For
>    I = 1...n+1, T'_{I} is initialized to the 128-bit string consisting
>    entirely of '0's. When an OP sends a relay cell, they prepare the
>    cell as follows:
>
>         The OP prepares the authentication part of the message:
>
>                 C_{n+1} = M
>                 T_{n+1} = Digest(Khf_n,T'_{n+1}||C_{n+1})
>                 N_{n+1} = T_{n+1} ^ E(Ktf_n,T_{n+1} ^ 0)
>                 T'_{n+1} = T_{n+1}
>
>         Then, the OP prepares the multi-layered encryption:
>
>                 For I=n...1:
>                         C_I = Encrypt(Kf_I,N_{I+1},C_{I+1})
>                         T_I = Digest(Khf_I,T'_I||C_I)
>                         N_I = T_I ^ E(Ktf_I,T_I ^ N_{I+1})
>                         T'_I = T_I
>
>           The OP sends C_1 and N_1 to node 1.
>
> 3.1.2. Relaying Forward at Onion Routers
>
>    When a forward relay cell is received by OR I, it decrypts the
>    payload with the stream cipher, as follows:
>
>         'Forward' relay cell:
>
>                 T_I = Digest(Khf_I,T'_I||C_I)
>                 N_{I+1} = T_I ^ D(Ktf_I,T_I ^ N_I)
>                 C_{I+1} = Decrypt(Kf_I,N_{I+1},C_I)
>                 T'_I = T_I
>
>    The OR then decides whether it recognizes the relay cell as
>    described below. If the OR recognizes the cell, it processes the
>    contents of the relay cell. Otherwise, it passes C_{I+1}||N_{I+1}
>    along the circuit if the circuit continues.
>
>    For more information, see section 4 below.
>
> 3.2. Backward Direction
>
>    The backward direction is the opposite direction from
>    CREATE/CREATE2 cells.
>
> 3.2.1. Relaying Backward at Onion Routers
>
>    When a backward relay cell is received by OR I, it encrypts the
>    payload with the stream cipher, as follows:
>
>         'Backward' relay cell:
>
>                 T_I = Digest(Khb_I,T'_I||C_{I+1})
>                 N_I = T_I ^ E(Ktb_I,T_I ^ N_{I+1})
>                 C_I = Encrypt(Kb_I,N_I,C_{I+1})
>                 T'_I = T_I
>
>    with C_{n+1} = M and N_{n+1}=0. Once encrypted, the node passes
>    C_I and N_I along the circuit towards the OP.
>
> 3.2.2. Routing to the Origin
>
>    When a relay cell arrives at an OP, the OP decrypts the payload
>    with the stream cipher as follows:
>
>         OP receives relay cell from node 1:
>
>                 For I=1...n, where n is the end node on the circuit:
>                         C_{I+1} = Decrypt(Kb_I,N_I,C_I)
>                         T_I = Digest(Khb_I,T'_I||C_{I+1})
>                         N_{I+1} = T_I ^ D(Ktb_I,T_I ^ N_I)
>                         T'_I = T_I
>
>                 If the payload is recognized (see Section 4.1),
>                 then:
>
>                        The sending node is I. Stop, process the
>                        payload and authenticate.
>
> 4. Application connections and stream management
>
> 4.1. Relay cells
>
>   Within a circuit, the OP and the end node use the contents of
>   RELAY packets to tunnel end-to-end commands and TCP connections
>   ("Streams") across circuits. End-to-end commands can be initiated
>   by either edge; streams are initiated by the OP.
>
>         The payload of each unencrypted RELAY cell consists of:
>
>                 Relay command           [1 byte]
>                 'Recognized'            [2 bytes]
>                 StreamID                [2 bytes]
>                 Length                  [2 bytes]
>                 Data                    [PAYLOAD_LEN-23 bytes]
>
>    The 'recognized' field is used as a simple indication that the
>    cell is still encrypted. It is an optimization to avoid
>    calculating expensive digests for every cell. When sending cells,
>    the unencrypted 'recognized' MUST be set to zero.
>
>    When receiving and decrypting cells the 'recognized' will always
>    be zero if we're the endpoint that the cell is destined for. For
>    cells that we should relay, the 'recognized' field will usually
>    be nonzero, but will accidentally be zero with P=2^-16.
>
>    If the cell is recognized, the node moves to verifying the
>    authenticity of the message as follows(*):
>
>           forward direction (executed by the end node):
>
>                 T_{n+1} = Digest(Khf_n,T'_{n+1}||C_{n+1})
>                 Tag = T_{n+1} ^ D(Ktf_n,T_{n+1} ^ N_{n+1})
>                 T'_{n+1} = T_{n+1}
>
>                 The message is authenticated (i.e., M = C_{n+1}) if
>                 and only if Tag = 0
>
>           backward direction (executed by the OP):
>
>                 The message is authenticated (i.e., C_{n+1} = M) if
>                 and only if N_{n+1} = 0
>
>
>    The old Digest field is removed since sufficient information for
>    authentication is now included in the nonce part of the payload.
>
>        (*) we should consider dropping the 'recognized' field
>        altogether and always try to authenticate. Note that this is
>        an optimization question and the crypto works just as well
>        either way.
>
>    The 'Length' field of a relay cell contains the number of bytes
>    in the relay payload which contain real payload data. The
>    remainder of the payload is padding bytes.
>
> 4.2. Appending the encrypted nonce and dealing with version-homogenic
>      and version-heterogenic circuits
>
>    When a cell is prepared to be routed from the origin (see Section
>    3.1.1) the encrypted nonce N is appended to the encrypted cell
>    (occupying the last 16 bytes of the cell). If the cell is
>    prepared to be sent to a node supporting the new protocol, S is
>    combined with other sources to generate the layer's
>    nonce. Otherwise, if the node only supports the old protocol, n
>    is still appended to the encrypted cell (so that following nodes
>    can still recover their nonce), but a synchronized nonce (as per
>    the old protocol) is used in CTR-mode.
>
>    When a cell is sent along the circuit in the 'backward'
>    direction, nodes supporting the new protocol always assume that
>    the last 16 bytes of the input are the nonce used by the previous
>    node, which they process as per Section 3.2.1. If the previous
>    node also supports the new protocol, these cells are indeed the
>    nonce. If the previous node only supports the old protocol, these
>    bytes are either encrypted padding bytes or encrypted data.
>
> 5. Security
>
> 5.1. Resistance to crypto-tagging attacks
>
>    A crypto-tagging attack involves a circuit with two colluding
>    nodes and at least one honest node between them. The attack works
>    when one node makes a change to the cell (tagging) in a way that
>    can be undone by the other colluding party. In between, the
>    tagged cell is processed by honest nodes which do not detect the
>    change. The attack is possible due to the malleability property
>    of CTR-mode: a change to a ciphertext bit effects only the
>    respective plaintext bit in a predicatble way. This proposal
>    frustrates the crypto-tagging attack by linking the nonce to the
>    encrypted message such that any change to the ciphertext results
>    in a random nonce and hence, random plaintext.
>
>    Let us consider the following 3-hop scenario: the entry and end
>    nodes are malicious and colluding and the middle node is honest.
>
> 5.1.1. forward direction
>
>    Suppose that node I tags the ciphertext part of the message
>    (C'_{I+1} != C_{I+1}) then forwards it to the next node (I+1). As
>    per Section 3.1.2. Node I+1 digests C'_{I+1} to generate T_{I+1}
>    and N_{I+2}. Since C'_{I+2} is different than it should be, so
>    are the resulting T_{I+1} and N_{I+2}. Hence, decrypting C'_{I+2}
>    using these values results in a random string for C_{I+2}. Since
>    C_{I+2} is now just a random string, it is decrypted into a
>    random string and cannot be 'recognized' nor
>    authenticated. Furthermore, since C'_{I+1} is different than what
>    it should be, T'_{I+1} (i.e., the running digest of the middle
>    node) is now out of sync with that of the OP, which means that
>    all future cells sent through this node will decrypt into garbage
>    (random strings).
>
>    Likewise, suppose that instead of tagging the ciphertext, Node I
>    node tags the encrypted nonce N'_{I+1} != N_{I+1}. Now, when Node
>    I+1 digests the payload the tweak T_{I+1} is find, but using it
>    to decrypt N'_{I+1} again results in a random nonce for
>    N_{I+2}. This random nonce is used to decrypt C_{I+1} into a
>    random C'_{I+2} which is not recognized by the end node. Since
>    C_{I+2} is now a random string, the running digest of the end
>    node is now out of sync, which prevents the end node from
>    decrypting further cells.
>
> 5.1.2. Backward direction
>
>    In the backward direction the tagging is done by Node I+2
>    untagging by the Node I. Suppose first that Node I+2 tags the
>    ciphertext C_{I+2} and sends it to Node I+1. As per Section
>    3.2.1, Node I+1 first digests C_{I+2} and uses the resulting
>    T_{I+1} to generate a nonce N_{I+1}. From this it is clear that
>    any change introduced by Node I+2 influences the entire payload
>    and cannot be removed by Node I.
>
>    Unlike in Section 5.1.1., the cell is blindly delivered by Node I
>    to the OP which decrypts it. However, since the payload leaving
>    the end node was modified, the message cannot be authenticated by
>    the OP which can be trusted to tear down the circuit.
>
>    Suppose now that tagging is done by Node I+2 to the nonce part of
>    the payload, i.e., N_{I+2}. Since this value is encrypted by Node
>    I+1 to generate its own nonce N_{I+1}, again, a random nonce is
>    used which affects the entire keystream of CTR-mode. The cell
>    again cannot be authenticated by the OP and the circuit is torn
>    down.
>
>    We note that the end node can modify the plain message before
>    ever encrypting it and this cannot be discovered by the Tor
>    protocol. This vulnerability is outside the scope of this
>    proposal and users should always use TLS to make sure that their
>    application data is encrypted before it enters the Tor network.
>
> 5.2. End-to-end authentication
>
>    Similar to the old protocol, this proposal only offers end-to-end
>    authentication rather than per-hop authentication. However,
>    unlike the old protocol, the ADL-construction is non-malleable
>    and hence, once a non-authentic message was processed by an
>    honest node supporting the new protocol, it is effectively
>    destroyed for all nodes further down the circuit. This is because
>    the nonce used to de/encrypt all messages is linked to (a digest
>    of) the payload data.
>
>    As a result, while honest nodes cannot detect non-authentic
>    messages, such nodes still destroy the message thus invalidating
>    its authentication tag when it is checked by edge nodes. As a
>    result, security against crypto-tagging attacks is ensured as
>    long as an honest node supporting the new protocol processes the
>    message between two dishonest ones.
>
> 5.3 The Running Digest
>
>    Unlike the old protocol, the running digest is now computed as
>    the output of a GHASH call instead of a hash function call
>    (SHA256). Since GHASH does not provide the same type of security
>    guarantees as SHA256, it is worth discussing why security is not
>    lost from computing the running digest differently.
>
>    The running digets is used to ensure that if the same payload is
>    encrypted twice, then the resulting ciphertext does not remain
>    the same. Therefore, all that is needed is that the digest should
>    repeat with low probability. GHASH is a universal hash function,
>    hence it gives such a guarantee assuming its key is chosen
>    uniformly at random.
> _______________________________________________
> tor-dev mailing list
> tor-dev@lists.torproject.org
> https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-dev