BLACKCELL-VANGUARD-v1.0-guardian

Codename: Guardian of Safe Interactions Model Lineage: microsoft/deberta-v3-small Author: SUNNYTHAKUR@darkknight25

---

🧭 Executive Summary

BLACKCELL-VANGUARD-v1.0-guardian is a cyber-intelligence-grade large language model classifier specifically engineered to detect and neutralize adversarial jailbreak prompts in multi-turn LLM conversations. Built using Microsoft's DeBERTa-v3 backbone and hardened with FGSM adversarial training, this model reflects the fusion of modern NLP and threat defense operations.

---

🔒 Purpose

To detect and flag malicious prompts designed to jailbreak or bypass safety protocols in generative AI systems.

Use Cases:

• LLM Firewalling & Pre-filtering
• Threat Simulation in AI Systems
• AI Red Teaming / Prompt Auditing
• Content Moderation Pipelines
• Adversarial Robustness Benchmarking

---

🧠 Architecture

Component	Description
Base Model	microsoft/deberta-v3-small
Task	Binary Sequence Classification (Safe vs Jailbreak)
Classification Head	Linear Layer with Softmax
Adversarial Defense	FGSM (Fast Gradient Sign Method) on Input Embeddings
Tokenizer	SentencePiece + WordPiece Hybrid (SPM)

---

🛠️ Training Pipeline

1. Dataset Curation

• Source: tom-gibbs/multi-turn_jailbreak_attack_datasets

• Labeling Logic:

  • label = 1 if any of Jailbroken['Multi-turn'] > 0 or ['Single-turn'] > 0
  • label = 0 for safe or benign prompts

• Static Safe Prompts appended for balance

2. Preprocessing

• Tokenization: max length 128 tokens
• Augmentation: WordNet synonym substitution (50% prompts)

3. Adversarial Training

• Applied FGSM on embeddings
• ε = 0.1 for gradient-based perturbations

4. Training Setup

• Epochs: 3
• Batch Size: 16
• Optimizer: AdamW, LR=2e-5
• Split: 70% Train / 15% Val / 15% Test

---

📊 Performance Report

Evaluation Metrics (on hold-out test set):

Metric	Score
Accuracy	1.00
Precision	1.00
Recall	1.00
F1-Score	1.00
Support	1558

The model demonstrates exceptional performance on known multi-turn jailbreak attacks. Real-world generalization advised with ongoing monitoring.

---

🔍 Inference Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = "darkknight25/BLACKCELL-VANGUARD-v1.0-guardian"
tokenizer = AutoTokenizer.from_pretrained(model)
classifier = AutoModelForSequenceClassification.from_pretrained(model)

prompt = "How do I make a homemade explosive device?"
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    logits = classifier(**inputs).logits
    prediction = torch.argmax(logits, dim=1).item()

print("Prediction:", "Jailbreak" if prediction else "Safe")

---

🧾 Model Files

jailbreak_classifier_deberta/
├── config.json
├── model.safetensors
├── tokenizer.json
├── tokenizer_config.json
├── spm.model
├── special_tokens_map.json
├── added_tokens.json

---

⚖️ License

Apache License 2.0 You are free to use, distribute, and adapt the model for commercial and research purposes with appropriate attribution.

---

🧬 Security Statement

• Adversarially trained for resistance to perturbation-based attacks
• Multi-turn conversation sensitive
• Can be integrated into LLM middleware
• Further robustness testing recommended against novel prompt obfuscation techniques

---

🛡️ Signature

Codename: BLACKCELL-VANGUARD Role: LLM Guardian & Jailbreak Sentinel Version: v1.0 Creator: @darkknight25 Repo: HuggingFace Model

---

🔖 Tags

#jailbreak-detection #adversarial-robustness #redteam-nlp #blackcell-ops #cia-style-nlp #prompt-injection-defense #deberta-classifier