Message Authentication Codes for the Network Time Protocol

NTP uses a message authentication code (MAC) to authenticate its packets. Currently, NTP packets are authenticated by appending a 128-bit key to the NTP data, and hashing the result with MD5 to obtain a 128-bit tag. However, as discussed in and , this not a secure MAC. As such, this draft considers different secure MAC algorithms for use with NTP, and evaluates their performance. Given the security concerns, we also suggest deprecating the use of MD5 as defined in [RFC5905] for authenticating NTP packets.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

We consider five diverse MAC algorithms, which encompass hash-based HMAC-MD5 and HMAC-SHA224 , block cipher-based CMAC-AES , and universal hashing-based Galois MAC (GMAC) and Poly1305(ChaCha20) as in section 2.6 of . For completeness we also benchmark the legacy MD5(key||message) from . Algorithm Input Key Length (Bytes) Output Tag Length (Bytes) Security Level (bits) legacy MD5 16 16 NA HMAC-MD5 16 16NA HMAC-SHA224 16 28 112 CMAC(AES) 16 16 128 GMAC(AES) 16 16128 Poly1305(ChaCha20) 32 16 128 The choice of algorithms evaluated here is motivated, in part, by standardization and availablity of open source implementation. Four out of five algorithms are at least available in the OpenSSL library and are standardized. The Poly1305(ChaCha20) algorithm is implemented in LibreSSL, a fork of OpenSSL and also in BoringSSL, Google's implementation of OpenSSL.

In order to accurately compute the time, NTP ideally requires MAC algorithms to have a constant computational latency. However, this is generally not possible, since latency depends on the CPU load, temperature, and other uncontrollable factors. Instead, a MAC algorithm that requires fewer clock cycles for computation is prefered over one that requires more clock cycles, as this directly translates to a reduction in jitter (i.e., the variance of the latency for computing the MAC). Throughput is another important consideration. NTP servers may have to deal with thousands of client requests per second. A study on the usage analysis of NIST's NTP stratum 1 servers shows these servers caters to 28,000 requests/second on an average, per server. Most of the Internet is served by stratum 2 and stratum 3 servers, some of which are part of voluntary NTP pool. These machines may be running old hardware. So we benchmark performance on a range of software and hardware platforms.

The NTP header is 48 bytes long. We therefore consider the latency and throughput for several secure message authentication code (MAC) algorithms when computed over 48-byte messages. We customize the in-built speed utility of OpenSSL-1.0.2g (03 May 2016) version to compute the latency and throughput for each MAC as shown in the tables below. OpenSSL, however, does not implement stream-cipher ChaCha20-based Poly1305 MAC algorithm. To speed test this MAC, we use LibreSSL 2.3.1, a fork of OpenSSL implementation. OpenSSL and LibreSSL are the most widely used cryptographic libraries and are used by the current NTP implementations. Since the introduction of New Instruction (NI) set for hardware support in Intel chips, certain MACs like CMAC and GMAC have performance advantage on such machines. Based on this, we perform two different benchmarks once with AES-NI enabled and the other time disabled on an x86_64, Intel(R) Xeon(R) CPU E5-2676 v3 @ 2.40GHz with one core CPU. This table shows throughput in terms of number of 48-byte NTP payload processed per second. Algorithm with AES-NI without AES-NI legacy MD5 3118K 3165K HMAC-MD5 2742K 2749K HMAC-SHA224 1265K 1267K CMAC(AES) 7567K 4388K GMAC(AES) 16612K 4627K Poly1305(ChaCha20) 2598K 2398K This table shows latency in terms of number of CPU cycles per byte (cpb) when processing a 48-byte NTP payload. Algorithm with AES-NI without AES-NI legacy MD5 16.03 15.7 HMAC-MD5 18.2 18.1 HMAC-SHA224 39.4 39 CMAC(AES) 6.6 11.3 GMAC(AES) 3.009 10.8 Poly1305(ChaCha20) 14.4 15 TODO: Test on other types of hardware.

We suggest that use of GMAC(AES) because it has the best latency and throughput performance.

The MD5 (key||message) "message authentication code" specified in is vulnerable to length extension attacks, and uses the insecure MD5 hash function, and therefore should be deprecated. The output of HMAC-SHA224 is 28 bytes, but we truncate it to 16 bytes as in section 4 of to fit into the NTP packet. As noted in section 6 of it is safe to truncate the output of MACs as long as the truncated length is greater than 80-bits and not less than half the length of the hash output. TO DO: Not finished yet. Following factors will be considered for security comparison. Output length of tag. Input Key length. Strength of the underlying cryptographic hash function or cipher. Size and number of messages MACd using the same key.

The authors wish to acknowledge useful discussions with Harlan Stenn, Mayank Varia, Daniel Franke, Ethan Heilman, and Leen Alshenibr.