fast-floe

An optimized, multi-provider, spec-complianct implementation of the Fast Lightweight Online Encryption (FLOE) scheme for authenticated encryption of large files and byte streams with bounded-memory and random access.

Introduction

fast-floe implements the Fast Lightweight Online Encryption (FLOE) specification in Rust. FLOE divides a message into independently authenticated and encrypted segments. A recipient can encrypt/decrypt a large message one segment at a time and read/seek randomly into the ciphertext.

FLOE is built for really big data, streaming data where you might not know size ahead of time, data that might arrive out-of-order, and random-access into encrypted data.

Callers can select a 4 KiB or 1 MiB segment size at encryption time to trade off throughput for random-access overhead (smaller segment size == cheaper random access, and larger segment size == more throughput).

Multiple cryptography providers are supported (aws-lc-rs, boring, ring, and RustCrypto). They all interoperate with each other (i.e. provider choice does not affect wire format).

Quick start

Add fast-floe from crates.io to Cargo.toml. Rust refers to the crate as fast_floe.

[dependencies]
fast-floe = "0.1"

aws-lc-rs is the default cryptographic provider, but can be overridden, see the "Cryptographic Providers" section below.

use fast_floe::{Key, Parameters, decrypt, encrypt};

# fn main() -> Result<(), fast_floe::Error> {
let key = Key::generate()?;
let aad = b"tenant=acme;object=backup";
let plaintext = b"data to protect";

let ciphertext = encrypt(
    &key,
    aad,
    Parameters::SEGMENT_4_KIB,
    plaintext,
)?;

let recovered = decrypt(&key, aad, &ciphertext)?;
assert_eq!(recovered, plaintext);
# Ok(())
# }

encrypt returns one complete FLOE message: an authenticated header followed by encrypted segments. decrypt authenticates the message before returning its plaintext.

The three inputs are:

• Key: a 32-byte secret. use Key::from_bytes to import existing key material and Key::generate to generate a new random key.
• AAD: context that must be authenticated with the ciphertext, such as an session ID or protocol version. FLOE does not store the AAD. Decryption requires the same AAD bytes to succeed.
• Parameters: the encrypted segment size. Choose 4 KiB for finer random reads or 1 MiB for higher throughput.

Choose an API

Ordered below from simplest (and more high-level) to more advanced (and low-level). Use the highest layer that fits your needs:

Whole messages

Use the quick-start encrypt and decrypt functions when the complete input and output fit in memory. Both return a new Vec<u8>.

decrypt uses the authenticated profile (segment size) read from the header. If your application wants to choose the profile instead, use decrypt_with_parameters.

Streams with std::io

EncryptReader and DecryptReader keep memory bounded while composing with ordinary Rust I/O:

use std::io::{self, Cursor};

use fast_floe::io::{DecryptReader, EncryptReader};
use fast_floe::{Key, Parameters};

# fn main() -> Result<(), Box<dyn std::error::Error>> {
let key = Key::from_bytes([0x42; Key::LEN]);
let aad = b"file metadata";

let input = Cursor::new(b"a potentially large input".as_slice());
let mut encrypting = EncryptReader::new(
    input,
    &key,
    aad,
    Parameters::SEGMENT_4_KIB,
)?;
let mut ciphertext = Vec::new();
io::copy(&mut encrypting, &mut ciphertext)?;

let mut decrypting = DecryptReader::new(Cursor::new(ciphertext), &key, aad)?;
let mut plaintext = Vec::new();
io::copy(&mut decrypting, &mut plaintext)?;
decrypting.finish()?;

assert_eq!(plaintext, b"a potentially large input");
# Ok(())
# }

Read EncryptReader to EOF to complete encryption. EncryptWriter is available when a Write interface fits better.

DecryptReader::finish authenticates unread ciphertext and rejects trailing bytes. Use finish_frame when non-FLOE data follows the FLOE message.

IMPORTANT: always call finish or try_finish to complete the FLOE message. flush will not emit a partial non-final FLOE segment, and dropping an unfinished writer leaves a truncated message.

Authenticated random access

random_access::Reader presents a seekable ciphertext as authenticated plaintext. It implements Read + Seek and also reads explicit ranges:

use std::io::Cursor;

use fast_floe::random_access::Reader;
use fast_floe::{Key, Parameters, encrypt};

# fn main() -> Result<(), Box<dyn std::error::Error>> {
let key = Key::from_bytes([0x42; Key::LEN]);
let aad = b"object 17";
let plaintext = vec![0x5a; 10_000];
let ciphertext = encrypt(
    &key,
    aad,
    Parameters::SEGMENT_4_KIB,
    &plaintext,
)?;

let mut reader = Reader::new(Cursor::new(ciphertext), &key, aad)?;
let range = reader.read_range(1_000..1_100)?;
assert_eq!(range, plaintext[1_000..1_100]);
# Ok(())
# }

Reader construction authenticates the header and final segment, which establishes the FLOE profile (segment size) and complete length. Each requested segment is authenticated before its bytes are returned. The underlying seekable source must remain unchanged while the reader is in use. Use Reader::new_with_length when non-FLOE data follows the last segment.

Segment-oriented processing

Use online::Encryptor and online::Decryptor when your transport already works with packets, buffers, or other segment-like things.

use std::io::{Cursor, Read};

use fast_floe::online::{Decryptor, Encryptor};
use fast_floe::{Key, Parameters};

# fn main() -> Result<(), Box<dyn std::error::Error>> {
let key = Key::generate()?;
let aad = b"example stream";
let parameters = Parameters::SEGMENT_4_KIB;
let segment_size = parameters.plaintext_segment_length();
let plaintext = vec![b'A'; segment_size * 2 + segment_size / 2];
let mut input = Cursor::new(&plaintext);

let mut encryptor = Encryptor::new(&key, aad, parameters)?;
let header = *encryptor.header();
let mut encrypted_segments = Vec::new();
let mut buffer = Vec::with_capacity(segment_size);

loop {
    buffer.clear();
    input
        .by_ref()
        .take(segment_size as u64)
        .read_to_end(&mut buffer)?;

    if buffer.len() < segment_size {
        encrypted_segments.push(encryptor.encrypt_final_segment(&buffer)?);
        break;
    }

    encrypted_segments.push(encryptor.encrypt_non_final_segment(&buffer)?);
}

let mut decryptor = Decryptor::new(&key, aad, &header)?;
let mut recovered = Vec::new();
for segment in encrypted_segments {
    recovered.extend(decryptor.decrypt_segment(&segment)?);
}
decryptor.finish()?;

assert_eq!(recovered, plaintext);
# Ok(())
# }

All non-final plaintext segments must have exactly parameters.plaintext_segment_length() bytes. The final segment may be shorter, including empty. Final encryption consumes the encryptor. Decryption reads the final marker from each segment, authenticates it, and finish detects a missing final segment (e.g. detects truncation). If the input ends exactly on a segment boundary, the next loop iteration emits an empty final segment.

Low-level API

fast_floe::low_level exposes the FLOE specification's details. A MessageLayout supplies the correct position, lengths, offsets, and finality for every segment and you should use it when you know the plaintext length in advance:

use fast_floe::low_level::start_encryption;
use fast_floe::{Key, Parameters};

# fn main() -> Result<(), Box<dyn std::error::Error>> {
let key = Key::from_bytes([0x42; Key::LEN]);
let aad = b"context stuff";
let parameters = Parameters::SEGMENT_4_KIB;
let plaintext = b"low-level api data example";
let layout = parameters.plaintext_layout(plaintext.len() as u64)?;

let (mut state, header) = start_encryption(&key, aad, parameters)?;
let mut ciphertext = header.as_ref().to_vec();

for segment in layout {
    let start = usize::try_from(segment.plaintext_offset())?;
    let end = start + segment.plaintext_length();
    let encrypted = state.encrypt_segment(&plaintext[start..end], segment)?;
    ciphertext.extend_from_slice(&encrypted);
}

assert_eq!(ciphertext.len() as u64, layout.ciphertext_length());
# Ok(())
# }

Parallel or concurrent workloads can call state.into_shared() and then shared.fork() for each distinct worker/thread. SharedEncryptionContext and SharedDecryptionContext are thread-safe (e.g. they are Send + Sync).

Low-level encryption requires the caller to handle all FLOE invariants: process every position once, produce exactly one final segment, leave no gaps, and process nothing after the final segment. Breaking these rules can break message security. Prefer the misuse-resistant higher-level APIs.

low_level::SegmentBuffer supports reusable in-place storage. If you need to control allocation, *_raw methods are available for use w/ your own buffer management.

Notes

• AAD is authenticated but is not stored in the ciphertext. Store or derive it separately and reproduce it exactly.
• Each segment is released only after its own authentication succeeds. A stream consumer may receive valid early segments before a later corruption or truncation is discovered. Callers should stage plaintext until finalization when the application requires all-or-nothing release.
• Segment prefixes and layout calculations describe framing. Treat them as untrusted until the corresponding header or segment authenticates.

Segment sizes

The crate supports two segment size choices:

• Parameters::SEGMENT_4_KIB
• Parameters::SEGMENT_1_MIB

Only the encrypted segment size differs, all other parameters (AES-256-GCM, HKDF-SHA-384, IV length) are the same.

Parameters::plaintext_layout and Parameters::ciphertext_layout calculate a complete random_access::MessageLayout. Use them for storage sizing and manual segment processing.

Cryptographic providers

This crate uses different providers to implement AES-256-GCM, HKDF-SHA-384, and random-number generation. The crate default is aws-lc-rs, but all of these are supported:

Provider choice does not change the FLOE wire format. They are all compatible with each other.

To use an alternate provider:

fast-floe = { version = "0.1", default-features = false, features = ["ring"] }

Provider features are additive. With exactly one compiled provider, Key::generate and Key::from_bytes use it automatically. But with several, you must bind one to the key with Key::generate_with_provider or Key::from_bytes_with_provider.

Examples and development

The repository includes two file examples:

• serial_file uses the bounded-memory std::io adapters.
• manual_file uses message layouts and low-level segment operations.

Run them with:

cargo run --example serial_file -- encrypt INPUT OUTPUT 64_HEX_KEY [4k|1m]
cargo run --example manual_file -- encrypt INPUT OUTPUT 64_HEX_KEY [4k|1m]

For a non-default provider, add --no-default-features --features ring before --.

Run the default-provider test and documentation checks with:

cargo test --all-targets
cargo test --doc
cargo clippy --all-targets -- -D warnings
RUSTDOCFLAGS="-D warnings" cargo doc --no-deps

Run all provider and segment-size benchmarks with:

./scripts/bench-matrix.sh

Benchmark results

Median throughput in GiB/s on an Apple M3 with each provider compiled as the sole provider. The "into" columns write to a separate output buffer while "in place" overwrite their input buffer.

License

Licensed under the Apache License 2.0.