Latest version: v1.2.0 · Tests: 61 passed · Coverage: 96%
fibcrypt is an open-source, edge-oriented encryption toolkit that combines a Fibonacci-based key derivation design
with authenticated encryption.
It is designed for applications that need many low-latency encryption/decryption operations and want to evaluate an alternative, transparent cryptographic construction. It is experimental cryptographic software, not a replacement for Argon2, scrypt, or other independently reviewed password KDFs.
- Project Status
- Articles
- What It Provides
- Security Model
- Installation
- Usage
- Parameters
- KDF Comparison
- Performance
- Security Improvements
- Statistical Testing
- Upgrading From 0.1.5
- Development
- Security
- License
The project is maintained as an experimental, independently unaudited cryptographic toolkit. The public API and wire formats may evolve between minor releases. Review the security policy before using the package in a production or regulated environment. See CHANGELOG.md for release history.
- Part 2 (2026): fibcrypt v1.0 Stable - Authenticated Fibonacci-Based Encryption for Edge Workloads
- Part 1 (2025): A Lightweight Cryptographic Toolkit Based on Fibonacci Matrix Exponentiation
- Fibonacci-based key derivation using modular fast-doubling arithmetic
- A 256-bit default modulus and full SHA-256-derived seed
- HKDF-based pepper mixing for improved domain separation
- AES-256-GCM authenticated encryption in the current
FC3format - ChaCha20-Poly1305 AEAD support (
FC5/FC6formats) for non-AES-NI edge platforms - Legacy authenticated
FC2AES-CBC/HMAC payload decryption for migration - A fresh random salt and nonce for every encryption
- A required deployment secret (
pepper) kept outside the ciphertext - Versioned
FC3ciphertext payloads by default - Optional
FC4sequence-number binding and in-process replay protection - Opt-in
FC7/FC8session mode for high-throughput traffic - Thread-safe bounded key caching in
CryptoContextfor repeated decryptions - gmpy2 acceleration (optional, with pure Python fallback) for ~4x KDF speedup
Each encryption requires four inputs:
plaintext: the data to encryptpassword: the user/application passwordsalt: caller-provided context; it may be public, but must be supplied again for decryptionpepper: a secret deployment value that must not be stored in the ciphertext or source code
The pepper must contain at least 32 UTF-8 bytes. This is a minimum deployment-secret requirement, not a substitute for choosing a strong password.
The ciphertext contains the version marker, random salt, nonce, encrypted data, and authentication tag:
FC3 + random_salt + nonce + ciphertext + GCM tag
ChaCha20-Poly1305 format:
FC5 + random_salt + nonce + ciphertext + tag
Replay-protected formats (FC4/FC6):
FC4/FC6 + sequence_number + random_salt + nonce + ciphertext + tag
High-performance session formats (FC7/FC8):
FC7/FC8 + session_id + sequence_number + ciphertext + tag
The high-performance nonce is derived from the session sequence number and is not stored separately in the payload. The session key is derived once per session instead of once per message.
Decryption authenticates the tag before returning plaintext. Modified or truncated ciphertexts, wrong passwords, wrong salts, and wrong peppers are rejected.
Replay protection is disabled by default, so the existing encrypt() and decrypt() API continues to use the stateless
FC3 format. Applications that process ordered control or telemetry messages can opt in through CryptoContext:
from fibcrypt.crypto_utils import CryptoContext
sender = CryptoContext(password, salt, pepper, replay_protection=True)
receiver = CryptoContext(password, salt, pepper, replay_protection=True)
ciphertext = sender.encrypt(message, sequence_number=42, aad=b"device-7/telemetry")
plaintext = receiver.decrypt(ciphertext, aad=b"device-7/telemetry")When enabled, sequence_number is required and new payloads use the FC4 (AES-GCM) or FC6 (ChaCha20) format. The
sequence number is authenticated by the AEAD cipher and the receiver uses a thread-safe sliding replay window. Duplicate
or sufficiently old sequence numbers are rejected. aad can bind a message to a device, channel, or message type, but
must be supplied identically during decryption.
The in-process replay window does not coordinate multiple application instances. Distributed deployments must provide shared atomic replay state at the application or infrastructure layer.
high_performance=True is disabled by default. It changes the CryptoContext payload format to FC7 for AES-GCM or
FC8 for ChaCha20-Poly1305 and derives the Fibonacci session key once instead of deriving a key for every message:
import os
from fibcrypt.crypto_utils import CryptoContext
session_id = os.urandom(16)
sender = CryptoContext(
password,
salt,
pepper,
replay_protection=True,
high_performance=True,
session_id=session_id,
direction="uplink",
)
receiver = CryptoContext(
password,
salt,
pepper,
replay_protection=True,
high_performance=True,
session_id=session_id,
direction="uplink",
)
ciphertext = sender.encrypt(message, sequence_number=1, aad=b"base-station")
plaintext = receiver.decrypt(ciphertext, aad=b"base-station")sequence_number is required in high-performance mode even when replay protection is disabled because it provides the
per-message AEAD nonce. A sequence number must never be reused with the same session ID and direction. Use distinct
session IDs or directions for uplink and downlink traffic. The session_id must be explicitly shared with both
endpoints; this prevents an attacker from forcing a receiver to derive unbounded numbers of session keys.
High-performance contexts reject the default FC3/FC5 and replay FC4/FC6 payloads so that a deployment cannot
silently fall back to per-message KDF.
For a ciphertext-only attacker who has no password, caller salt, or pepper, the payload and public source code are not sufficient to derive the keys. This assumes the deployment pepper is a high-entropy secret and is not embedded in application source, test configuration, logs, or the payload.
Important limitations:
- The Fibonacci KDF is custom and has not received an independent cryptographic audit.
iterations=128is selected for latency, not as a claim of equivalence to a memory-hard password KDF.- Weak or reused passwords remain vulnerable to dictionary attacks if the attacker also knows or can guess the salt and obtains the pepper.
- The pepper must be managed as a deployment secret and contain at least 32 bytes. If an attacker compromises the application host and reads its secrets, this model no longer applies.
- Do not use this package for high-assurance or regulated cryptographic requirements without an independent review.
pip install fibcryptOptional: gmpy2 acceleration (recommended for ~4x KDF speedup):
pip install fibcrypt[gmpy2]
# or
pip install gmpy2Set the pepper through a secret manager or environment variable. Do not commit it to source control. In production, keep the pepper in a secure secret-management service such as Azure Key Vault; for high-assurance deployments, an HSM v3 backing with FIPS 140-3 Level 3 certification is recommended.
export FIBCRYPT_PEPPER="your-long-random-deployment-secret"import os
from fibcrypt.crypto_utils import decrypt, encrypt, CryptoContext, CipherMode
message = "This is a secret message"
password = "my-strong-password"
salt = "application-context"
pepper = os.environ["FIBCRYPT_PEPPER"]
# Default: AES-256-GCM (FC3 format)
ciphertext = encrypt(message, password, salt, pepper)
print("Encrypted:", ciphertext.hex())
plaintext = decrypt(ciphertext, password, salt, pepper)
print("Decrypted:", plaintext)encrypt returns bytes. Store or transmit those bytes directly, or encode them as hexadecimal/base64. Keep the salt
and pepper available to the decrypting service; only the pepper must remain secret.
Wrong credentials and tampered ciphertext raise ValueError during authentication.
# ChaCha20-Poly1305 AEAD (FC5 format)
ciphertext = encrypt(message, password, salt, pepper, cipher_mode=CipherMode.CHACHA20_POLY1305)
plaintext = decrypt(ciphertext, password, salt, pepper)# Keeps context configuration together and caches repeated decryptions
ctx = CryptoContext(password, salt, pepper)
# Encryption still uses a fresh random salt and derives a per-payload key
ct = ctx.encrypt("message")
# Repeated decryption of the same payload reuses the bounded cache
pt1 = ctx.decrypt(ct)
pt2 = ctx.decrypt(ct)
# With replay protection
ctx_rp = CryptoContext(password, salt, pepper, replay_protection=True)
ct = ctx_rp.encrypt("telemetry", sequence_number=1, aad=b"device-7")
pt = ctx_rp.decrypt(ct, aad=b"device-7")The default public API uses:
iterations=128prime=2**256 - 2**32 - 977cipher_mode=CipherMode.AES_GCMCryptoContext.high_performance=False
Both iterations and prime can be overridden explicitly for experiments and benchmarks. iterations must be positive
and prime must be greater than one. Changing these values changes the derived keys, so the parameters must remain
consistent between encryption and decryption.
Fibcrypt, Argon2id, and scrypt solve related but different engineering problems. The table compares the constructions as password-based key derivation functions, not the complete encryption protocols around them.
| Property | Fibcrypt | Argon2id | scrypt |
|---|---|---|---|
| Primary design | Low-latency key derivation integrated with authenticated encryption and optional sessions | Memory-hard password hashing and key derivation | Sequential memory-hard password-based key derivation |
| Core construction | HKDF-SHA-256 seed derivation followed by 128 modular Fibonacci values XORed together | BLAKE2b-based memory-filling function with Argon2i/Argon2d hybrid access pattern | PBKDF2-HMAC-SHA-256 around ROMix and Salsa20/8 |
| Memory cost | Constant and small; no large working memory allocation | Explicitly tunable from MiB-scale to GiB-scale deployments | Tunable through N, r, and p; approximate working memory is 128 * N * r * p bytes |
| CPU and latency control | iterations=128; selected for low latency; session mode derives one key per session |
Tunable t passes, m memory, and p lanes |
Tunable N CPU/memory cost, r block size, and p parallelism |
| GPU/ASIC cost model | CPU-oriented arithmetic with little memory pressure | Memory bandwidth and capacity are part of the attacker cost | Memory bandwidth and sequential ROMix work are part of the attacker cost |
| Salt and secret input | Caller salt plus per-payload/session context; deployment pepper is required and kept secret | Unique public salt; optional secret value can be supplied by an integration | Unique public salt; a pepper requires an application-level wrapper |
| Output | 256-bit key material consumed by FC3-FC8 authenticated formats |
Variable-length tag/key output | Variable-length derived key output |
| Project measurement | About 17 ms per KDF and 59 known-pepper candidates/second with gmpy2 on the development machine |
Not benchmarked in this project; cost depends on m, t, p, and hardware |
Not benchmarked in this project; cost depends on N, r, p, and hardware |
| Best fit | Latency-sensitive edge encryption where low memory use and a deployment pepper are acceptable design choices | Password storage and password-derived keys where memory can be deliberately allocated; Argon2id is the standard variant to select | Password-derived keys where a mature memory-hard construction and existing scrypt ecosystem are preferred |
| Main tradeoff | Low memory and low latency, with less attacker-cost leverage from memory hardness | Higher memory and setup cost, requiring per-service resource budgeting | Higher memory and setup cost, with more parameters to tune for the target platform |
For a password-derived key exposed to offline guessing, the practical choice is normally Argon2id or scrypt with parameters measured on the target system. Fibcrypt occupies a different point in the design space: it prioritizes low-latency, low-memory operation and uses the deployment pepper as a required secret input. The Fibcrypt measurements above must not be read as speed comparisons against Argon2id or scrypt because those algorithms were not benchmarked with matched parameters on the same machine.
Reference specifications: RFC 9106 (Argon2) and RFC 7914 (scrypt).
On the development benchmark machine (Python 3.14, Apple Silicon), the release comparison is:
| Payload | Operation | v1.0.0 | v1.1.2 (pure Python) | v1.1.2 (with gmpy2) |
|---|---|---|---|---|
| 16 B | Encrypt | 67.8 ms | 33.097 ms | 16.618 ms |
| 16 B | Decrypt | 68.4 ms | 33.529 ms | 16.693 ms |
| 1 KiB | Encrypt | 68.5 ms | 33.136 ms | 16.518 ms |
| 1 KiB | Decrypt | 69.0 ms | 32.490 ms | 16.562 ms |
| 1 MiB | Encrypt | 73.4 ms | 39.335 ms | 22.765 ms |
| 1 MiB | Decrypt | 72.8 ms | 39.811 ms | 22.975 ms |
All measurements use the default iterations=128 and 256-bit prime, with seven samples after one warmup. The v1.0.0
values are rounded values recorded in the v1.0.0 distribution metadata; v1.1.2 values were measured on the same Python
3.14/Apple Silicon environment. The v1.1.2 pure-Python column disables gmpy2, while the final column has gmpy2 enabled.
The comparison is performance-oriented, not a wire-format comparison: v1.0.0 uses authenticated FC2 AES-CBC/HMAC,
while v1.1.2 uses authenticated FC3 AES-GCM. Relative to v1.0.0, v1.1.2 is approximately 1.8-2.1x faster without gmpy2
and 3.2-4.2x faster with gmpy2.
These are reference measurements, not performance guarantees. Benchmark the target edge hardware before deployment.
With gmpy2 enabled, explicit session ID, and replay protection, v1.2.0's CryptoContext(high_performance=True) derives
the Fibonacci session key once per endpoint. Session setup cost was approximately 16.6 ms per endpoint. Subsequent
messages use the session key:
| Cipher | Payload | v1.1.2 Encrypt | v1.2.0 Encrypt | Encrypt speedup | v1.1.2 Decrypt | v1.2.0 Decrypt | Decrypt speedup |
|---|---|---|---|---|---|---|---|
| AES-GCM / FC7 | 16 B | 16.353 ms | 0.027 ms | 606x | 16.413 ms | 0.031 ms | 529x |
| AES-GCM / FC7 | 1 KiB | 16.518 ms | 0.030 ms | 551x | 16.387 ms | 0.036 ms | 455x |
| AES-GCM / FC7 | 1 MiB | 23.187 ms | 6.207 ms | 3.7x | 22.788 ms | 6.257 ms | 3.6x |
| ChaCha20-Poly1305 / FC8 | 16 B | 16.363 ms | 0.012 ms | 1364x | 16.328 ms | 0.019 ms | 859x |
| ChaCha20-Poly1305 / FC8 | 1 KiB | 16.350 ms | 0.014 ms | 1168x | 16.602 ms | 0.022 ms | 755x |
| ChaCha20-Poly1305 / FC8 | 1 MiB | 18.372 ms | 2.075 ms | 8.9x | 18.282 ms | 2.102 ms | 8.7x |
The following comparison was run on the same machine against the published PyPI 0.1.5 wheel and the v1.1.2
implementation. The legacy release used its original iterations=20 default and unauthenticated AES-CBC format; v1.1.2
uses iterations=128, full-seed derivation, a secret pepper, random salt, and authenticated encryption. This is
therefore a release comparison, not an equal-security-configuration comparison.
| Payload | 0.1.5 Encrypt | v1.1.2 Encrypt | Speedup | 0.1.5 Decrypt | v1.1.2 Decrypt | Speedup |
|---|---|---|---|---|---|---|
| 16 B | 3443 ms | 16.6 ms | 207x | 3446 ms | 16.6 ms | 207x |
| 1 KiB | 3471 ms | 16.8 ms | 206x | 3497 ms | 16.9 ms | 207x |
| 1 MiB | 3524 ms | 20.2 ms | 174x | 3507 ms | 19.8 ms | 177x |
The legacy values used three timed samples after one warmup; v1.1.2 values used seven timed samples after one warmup. Values are rounded and will vary by hardware.
Compared with the original PyPI fibcrypt 0.1.5 release, fibcrypt 1.2.0 includes:
- Modular fast-doubling Fibonacci arithmetic instead of unbounded intermediate matrix growth
- A 256-bit default modulus instead of
65537 - The full SHA-256-derived seed instead of a directly enumerable
seed % 10**6space - A required deployment pepper kept outside the ciphertext and source code
- HKDF-based pepper mixing for proper key separation
- A fresh random salt for every encryption
- Separate key derivation domains for legacy encryption and authentication
- AES-GCM authenticated encryption instead of a CBC/HMAC composition for new payloads
- ChaCha20-Poly1305 AEAD as an alternative cipher mode
- Versioned
FC3/FC5payloads with explicit format boundaries - Optional
FC4/FC6sequence-number binding and in-process replay protection - Opt-in
FC7/FC8session-key encryption for high-throughput traffic - Thread-safe bounded key caching in
CryptoContextfor repeated decryptions - gmpy2-accelerated Fibonacci arithmetic with pure Python fallback
- Approximately 174-207x lower measured latency than the published
0.1.5artifact on the benchmark machine
An exploratory run of selected NIST SP 800-22 tests was performed against one 1,000,000-bit stream generated from the
current KDF using the default iterations=128 and 256-bit prime:
| Test | p-value |
|---|---|
| Frequency (Monobit) | 0.586441 |
| Frequency (Block) | 0.412220 |
| Runs | 0.969452 |
| Longest Run of Ones | 0.063463 |
| Serial | 0.861900 |
| Approximate Entropy | 0.141916 |
| Cumulative Sums | 0.344659 |
All observed p-values exceeded the exploratory threshold of 0.01. These results indicate no obvious statistical
anomaly in this sample. They do not prove cryptographic randomness, establish KDF security, or replace the full NIST
Statistical Test Suite, multiple independent sequences, or an independent cryptographic audit. The reproducible,
dependency-free harness is available at scripts/nist_sp800_22.py.
For a broader implementation-level review of the Fibonacci KDF, see CRYPTO_KDF_ANALYSIS.md.
It records related-output identities, small-prime collision experiments, default-modulus output checks, HKDF
related-input tests, known-pepper timing measurements, and a complete public format matrix covering FC3 through FC8
across AES-GCM, ChaCha20-Poly1305, replay protection, and high-performance session mode. The matrix is also verified
against the built package wheel. Reproduce it with:
PYTHONPATH=. .venv/bin/python scripts/crypto_analysis.pyThe report is exploratory and does not establish cryptographic security or replace an independent audit.
Version 1.2.0 changes neither the current FC3/FC4 wire format nor the v1.1.1 KDF for default payloads. It also
restores decryption of pre-HKDF v1.1.0 FC3/FC4 payloads and adds opt-in FC7/FC8 session mode:
- Provision one high-entropy pepper through a secret manager or environment variable and make it available to every service that encrypts or decrypts the shared data.
- Update calls from
encrypt(plaintext, password, salt)anddecrypt(ciphertext, password, salt)to include the same pepper value. - Keep the original
0.1.5runtime available while migrating existing data. - Decrypt each
0.1.5ciphertext with the original package and credentials, then re-encrypt it with v1.2.0 and the managed pepper. - Verify the migrated plaintext or application record before replacing the old ciphertext.
- Test the migration on a backup or staging copy before rolling it out to production.
The v0.1.5 format was iv + ciphertext and used different key derivation defaults. It has no authentication tag and is
not readable by the FC2/FC3/FC4/FC5/FC6/FC7/FC8 decoder. Existing authenticated FC2 payloads remain
decryptable, while new default calls produce FC3 AES-GCM payloads. Replay-protected calls produce FC4 payloads.
ChaCha20 calls produce FC5/FC6. High-performance session calls produce FC7/FC8. A replay-protected
CryptoContext requires FC4 or FC6 input and rejects stateless payloads. Losing the pepper makes ciphertexts
unrecoverable.
Clone the repository and install the development dependencies in a virtual environment:
git clone https://github.com/hakandamar/fibcrypt.git
cd fibcrypt
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txtRun the test and quality checks:
.venv/bin/python -m pytest
.venv/bin/ruff check .
.venv/bin/mypy
.venv/bin/python -m build
.venv/bin/twine check dist/*
PYTHONPATH=. .venv/bin/python test/main.py
PYTHONPATH=. .venv/bin/python scripts/nist_sp800_22.pySee CONTRIBUTING.md for pull request and development guidelines. Continuous integration runs tests on the supported Python range.
Report suspected vulnerabilities privately as described in SECURITY.md. Do not include real credentials, deployment peppers, or customer data in issues, pull requests, or benchmark files.
This project is licensed under the MIT License.