@ -46,7 +46,7 @@ The padding scheme is the same as used with Twofish.
AES relies on openSSL's `evp_*` interface which also offers hardware acceleration where available (SSE, AES-NI, …). It however is slower than the following stream ciphers because the CBC mode cannot compete to the optimized stream ciphers; maybe AES-CTR being a stream cipher could.
AES relies on openSSL's `evp_*` interface which also offers hardware acceleration where available (SSE, AES-NI, …). It however is slower than the following stream ciphers because the CBC mode cannot compete to the optimized stream ciphers; maybe AES-CTR being a stream cipher could.
_Current ideas are to bring CTS mode to AES in some future version, just to avoid unneccessary weight gains from padding. CTS mode works well starting with plain texts from one block plus one byte. So, we might revert back to the Twofish-way of IV handling._
_Current ideas are to bring CTS mode to AES in some future version, just to avoid unneccessary weight gains from padding. CTS mode works well starting with plain texts from one block plus. So, we might revert back to the Twofish-way of IV handling with a full block IV._
### ChaCha20
### ChaCha20
@ -62,16 +62,139 @@ ChaCha20 usually performs faster than AES-CBC.
SPECK is recommend by the NSA for offical use in case AES implementation is not feasible due to system constraints (performance, size, …). The block cipher is used in CTR mode making it a stream cipher. The random full 128-bit IV is transmitted in plain.
SPECK is recommend by the NSA for offical use in case AES implementation is not feasible due to system constraints (performance, size, …). The block cipher is used in CTR mode making it a stream cipher. The random full 128-bit IV is transmitted in plain.
On Intel CPUs, SPECK performs even faster than openSSL's ChaCha20 as it takes advantage of SSE4 or AVX2 if available. Though, on Raspberry's ARM CPU, it is second place behind ChaCha20 and before AES-CBC.
On Intel CPUs, SPECK performs even faster than openSSL's ChaCha20 as it takes advantage of SSE4 or AVX2 if available (compile using `-march=native`). On Raspberry's ARM CPU, it is second place behind ChaCha20 and before AES-CBC.
_An ARM specific optimized implementation (NEON?) is still missing. Also, multi-threading might accelerate this cipher on all CPUs with more than one core._
_An ARM specific optimized implementation (NEON?) is still missing. Also, multi-threading might accelerate this cipher on all CPUs with more than one core._
### Random Numbers
### Random Numbers
Throughout n2n, pseudo-random numbers are generated for several purposes, e.g. random MAC assignment and the IVs for use with the various ciphers. With a view to the IVs, especially for use in the stream ciphers, the pseudo-random numbers shall be as collision-free as possible. n2n uses an implementation of XORSHIFT128+ which shows a periodicity of 2¹²⁸.
Its initialization relies on seeding with a value as random as possible. Various sources are tapped including a syscall to Linux' `SYS_getrandom` as well as Intels hardware random number generators `RDRND` and `RDSEED`, if available (compile using `-march=native`).
### Pearson Hashing
For general purpose hashing, n2n employs Pearson hashing as it offers different hash sizes and is said to not be too collidy. However, this is not a cryptographically secure hashing function which by the way is not required here: The hashing is never applied in a way that the hash shall proove the knowledge of a secret without showing the secret.
_Pearson hashing is tweakable by making your own permutation of the 256 byte table._
_Pearson hashing allows for verifying only parts of the hash – just in case performance requirements would urge to do so._
## Header
## Header
### Overview
Packet's header consist of a COMMON section followed by a packet-type specific section, e.g. REGISTER, REGISTER_ACK, PACKET including the payload, REGISTER_SUPER, …
If enabled (`-H`), all fields but the Payload (which is handled seperately as outlined above) get encrypted using SPECK in CTR mode. As packet headers need to be decryptable by the supernode and we do not want to add another key (keep it a simple interface), the community name serves as key (keep it secret!) because it is already known to the supernode.
The scheme applied tries to maintain compatibility with current packet format and works as follows:
- First line of 4 bytes (Version, TTL, Flags) goes to fifth line:
- To later be able to identify a correctly decrpyted header, a magic number is stamped in fourth line starting at byte number 12. We use "n2n" string and add the header length to later be able to stop header decryption right before an eventually following payload begins – in case of PACKET-type, header-length does not equal packet-length.
- The rest of the community field, namely the first 12 bytes, is reframed towards a 96-bit IV for the header encryption.
- As we use a stream cipher, the IV should be a nonce. The IV plays an additional role sketched later, see the following sections on checksum and replay protection.
- To make a less predictable use of the key space – just think of usually reset MSB of ASCII characters of cimmunity names – we actually use a hash of the community name as key.
Decrpytion checks all known communities (several in case of supernode, only one at edge) as keys. On success, the emerging magic number will reveal the correct community whose name will be copied back to the original fiels allowing for regular packet handling.
Thus, header encryption will only work with previously determined community names introduced to the supernode by `-c <path>` parameter. Also it should be clear that header encryption is a per-community decision, i.e. all nodes and the supernode need to have it enabled. However, the supernode supports encrpyted and unencrypted communities in parallel, it determines their status online at arrival of the first packet. Use a fresh community name for encrypted communities, do not use a previously used one as in unecrpyted communities, the names were openly transmitted.
### Checksum
### Checksum
The whole packet including the eventually present payload is checksummed using a modified Person hashing. It might seem a little short compared to usual message tags of 96 up to 128 bit, especially when using a stream cipher which easily allows for bit-flips. So, the 16-bit checksum is filled up with 80 more bits to obtain 96-bit pre-IV. This pre-IV gets encrypted using a single block-cipher step to get the pseudo-random looking IV. This way, the checksum resists targeted bit-flips as any change to the whole 96-bit IV would render the header un-decryptable.
The single block-cipher step employs SPECK as it is fast and offers a 96-bit version, the key is derived from the header key – a hash of the hash.
The checksum gets verified by the edges and the supernode.
### Replay Protection
### Replay Protection
The aforementioned fill-up does not completely rely on random bits. A 52-bit time stamp displaying a microsecond-accuracy is encoded to the 96-bit pre-IV as well:
Encrypting this pre-IV with a block cipher step will generate a pseudo-random looking IV which gets written to the packet and used for the header encryption.
Due to the time-stamp encoded, the IV will more likely be unique, e.g. almost assuredly be a nonce.
Upon receival, the time stamp as well as the checksum can be extracted from the IV by performing a 96-bit block-cipher decryption step. Verification of the time stamp happens in two steps:
- The time stamp is checked against the local clock. It may not deviate more than plus/minus 16 seconds. So, edge and supernode need to keep a somewhat current time. This limit can be adjusted by changing the `TIME_STAMP_FRAME` definition. It is time-zone indifferent as UTC is used.
- Valid time stamps get stored as "last valid time stamp" seen from each node (supernode and edges). So, a newly arriving packet's time stamp can be compared to the last valid one. It should be equal or higher. However, as UDP packets may overtake each other just by taking another path through the internet, they are allowed to be 160 millisecond earlier than the last valid one. This limit can be adjusted by changing the `TIME_STAMP_JITTER` definition.
Logan007
4 years ago