SamSec007/phishbyte

Phish_Byte v7

A from-scratch PyTorch model for email phishing detection.

F1 0.950 on 5,000 held-out samples from a 6-corpus benchmark. 254K parameters (≈260× smaller than DistilBERT). 995 emails/sec on a laptop GPU. 85 engineered features (35 rule-based + 50 TF-IDF learned from corpus). Every verdict explains itself with full per-feature attribution.

The only non-transformer phishing detection model on HuggingFace.

Quick start

from phishbyte import PhishByteEngine

engine  = PhishByteEngine.from_pretrained("SamSec007/phishbyte")
verdict = engine.analyze(raw_email_string)

print(verdict.label)             # "phishing"
print(verdict.probability)       # 0.9735
print(verdict.confidence)        # "high"
print(verdict.layer_used)        # 2
print(verdict.feature_weights)   # 85-feature attribution

Analyse a real email from Gmail

1. Open the email in Gmail
2. Click ⋮ → Show original
3. Copy all (Ctrl+A, Ctrl+C)

engine = PhishByteEngine.from_pretrained("SamSec007/phishbyte")
verdict = engine.analyze(pasted_raw_email)
print(verdict)

Or save as .eml and run:

python cli.py --file suspicious.eml

What changed in v7

• 85 features (was 29) — added 50 TF-IDF unigrams + 3 BDI features + 2 domain features + 1 composite
• 254K parameters (was 12K) — deeper residual MLP with two ResBlocks and input skip connection
• 6-dataset training (was CEAS-2008 only) — Enron, SpamAssassin, Ling-Spam, Nazario, Nigerian Fraud
• TF-IDF vocabulary — 50 most discriminative unigrams learned from training corpus. No pretrained LM.
• Body Domain Identification — most common link domain mismatch, form action mismatch, external link ratio
• Display name spoofing — catches "PayPal Security" <attacker@evil.com>
• Calibrated training metrics — F1 at Youden-optimal threshold, not naive 0.5 cutoff

Architecture

raw email
  → Layer 1 (6 rule scorers, ~1ms) → veto gate (obvious phishing only)
  → Layer 2 (residual MLP, ~3ms)
      85 → 360 → 180 (×2 ResBlock) → 90 → 48 → 1 (sigmoid)
      + input-to-output skip connection
  → PhishVerdict {label, probability, confidence, layer_used, feature_weights}

Benchmarks (5,000 held-out, 6-corpus)

Metric	Phish_Byte v7	DistilBERT fine-tuned
F1 score	0.950	~0.967
Accuracy	94.94%	~97%
Parameters	254K	66,000,000
Model size	~1 MB	~263 MB
Throughput (GPU)	995/sec	~50/sec
GPU required	No	Practically yes
Header + SPF analysis	Yes	No
Per-feature attribution	85 features	Token-level SHAP

Feature groups (85 total)

Group	Count	Examples
Domain	7	mismatch, Reply-To diff, brand impersonation, display name spoof, suspicious pattern
URL + Body	10	HTTPS ratio, anchor mismatch, urgency (normalized), caps ratio, digit ratio
SPF	3	fail, no record, no IP
Subject	7	urgency, security theme, brand, currency, all caps, fake RE, fake txn ID
BDI	3	most common link domain mismatch, form action mismatch, external link ratio
TF-IDF	50	top-50 discriminative unigrams from training corpus
Composite	5	per-module layer scores

Training data

CEAS-2008 + Enron + SpamAssassin + Ling-Spam + Nazario + Nigerian Fraud = ~83K emails (balanced 50/50).

Same 6-corpus benchmark used by the top DistilBERT model on HuggingFace.

Install

pip install huggingface_hub safetensors dnspython

Limitations

• ~5% error rate. Use as one signal in defence-in-depth.
• Trained on English-language phishing (2003–2008 era). Modern attacks and non-English emails will degrade recall.
• SPF validation skipped for training (historical domains). Re-enables at inference on live emails.
• TF-IDF vocabulary is corpus-specific. Retrain on your own data for best domain fit.

Citation

@software{phishbyte2026,
  author = {Singh, Samratth},
  title  = {Phish_Byte: Cascading from-scratch PyTorch phishing detection},
  year   = {2026},
  url    = {https://github.com/AnonymousSingh-007/Phish_Byte}
}

License

MIT