CogPIE: Cognitive Pattern Inference Engine

A hybrid multi-layer password generation engine that exploits cognitive biases in human password creation. CogPIE combines semantic expansion, cognitive transforms, Markov+PCFG hybrid ranking, entropy scoring, and policy filtering to predict PII-derived passwords with high accuracy.

Architecture

CogPIE implements a 5-layer pipeline:

1. Semantic Expander: Generates base candidates from profile data (names, dates, keywords, pets, companies)
2. Cognitive Transforms: Applies psychologically-motivated transformations (leetspeak, case variations, suffix patterns, anchoring bias)
3. Markov+PCFG Hybrid Ranker: Scores candidates using a trained character-level Markov model combined with PCFG structural templates
4. Entropy Scorer: Filters and re-ranks based on entropy thresholds
5. Policy Filter: Applies password policy constraints (length, complexity requirements)

Key Results

Comparative Analysis (2,000 targets, 1M training)

CogPIE surpasses CUPP at budget >= 1,000 guesses. Markov collapses catastrophically at scale (28.5% -> 4.8%).

GPU Acceleration

CUDA backend with 20x speedup on RTX 4060 (1M corpus):

• CPU: 15.12s/profile
• GPU: 0.75s/profile (174 MB VRAM)

Installation

pip install cogpie

For GPU acceleration (CUDA):

pip install cogpie[gpu]

CLI

cogpie --profile target.json --corpus rockyou.txt --max 50000 -o candidates.txt

Usage

from cogpie.engine import CogPIEEngine

engine = CogPIEEngine(use_gpu=True)

# Train on a password corpus
with open("corpus.txt") as f:
    for line in f:
        engine.ranker.train(line.strip())

# Generate candidates for a profile
profile = {
    "names": ["joao"],
    "dates": ["1990", "0315"],
    "keywords": ["flamengo"],
    "pets": ["rex"],
    "companies": ["acme"],
    "phones": [],
    "emails": ["joao@example.com"],
    "old_passwords": ["joao123"],
    "numbers": ["42"],
}
candidates = engine.generate(profile, max_candidates=10000)
for password, score in candidates[:10]:
    print(f"{password}: {score:.4f}")

Research Paper

The full academic paper is available in docs/research/cogpie-paper/:

• English: cogpie-paper.md / cogpie-paper.html
• Portuguese (BR): cogpie-paper-pt-br.md / cogpie-paper-pt-br.html

42 references, MITRE ATT&CK/CWE/OWASP/NIST D3FEND mappings, 55 unit tests as experimental evidence.

Requirements

• Python >= 3.10
• PyTorch >= 2.0 (optional, for GPU acceleration)
• CUDA-capable GPU (optional, falls back to CPU)

Author

Andre Henrique de Oliveira (@mrhenrike) / Uniao Geek

License

MIT License. See LICENSE for details.