cyberguard-ai-security-analyzer

🔒 CyberGuard AI - Security Log Analyzer

Model Description

CyberGuard AI is a fine-tuned language model specialized in analyzing cybersecurity logs and identifying potential security threats. Built on Llama-2-7B and trained using Unsloth for efficient fine-tuning.

Developed by: omasteam

🎯 Model Capabilities

• Threat Classification: Categorizes logs as Normal, Suspicious, or Malicious
• Confidence Scoring: Provides confidence levels for classifications
• Detailed Analysis: Offers comprehensive reasoning for each assessment
• Actionable Recommendations: Suggests appropriate security responses

📊 Classification Categories

🟢 Normal

• Routine system operations
• Standard user activities
• Expected network traffic
• Scheduled maintenance events

🟡 Suspicious

• Unusual access patterns
• Failed authentication attempts
• After-hours activities
• Anomalous data transfers

🔴 Malicious

• Active attack indicators
• Malware signatures
• SQL injection attempts
• Data exfiltration patterns

🚀 Quick Start

from unsloth import FastLanguageModel

# Load the model
model, tokenizer = FastLanguageModel.from_pretrained(
    "{omasteam}/{cyberguard-ai-security-analyzer}",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Set to inference mode
FastLanguageModel.for_inference(model)

# Analyze a security log
messages = [
    {{
        "role": "system",
        "content": "You are CyberGuard AI, an expert cybersecurity analyst. Analyze security logs and classify them as 'normal', 'suspicious', or 'malicious'."
    }},
    {{
        "role": "user",
        "content": "Analyze this security log: Multiple failed login attempts from IP 192.168.1.100"
    }}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt"
)

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=256,
    temperature=0.3,
    do_sample=True
)

response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)
print(response)

🔧 Training Details

Training Data

• Dataset Size: 150 security log examples
• Categories: Normal (50), Suspicious (50), Malicious (50)
• Format: Conversational training with system prompts

Training Configuration

• Base Model: Llama-2-7B-Chat
• Training Method: LoRA (Low-Rank Adaptation)
• Training Steps: 60
• Learning Rate: 2e-4
• Batch Size: 2 (with gradient accumulation)

Training Framework

• Framework: Unsloth (optimized training)
• Quantization: 4-bit for memory efficiency
• Hardware: Google Colab GPU

📈 Performance

The model demonstrates strong performance in:

• Threat Detection: High accuracy in identifying malicious activities
• False Positive Reduction: Balanced classification to minimize alerts
• Contextual Understanding: Considers temporal and behavioral patterns
• Response Quality: Provides detailed, actionable analysis

⚠️ Limitations

• Trained on synthetic security log data
• May require fine-tuning for specific environments
• Performance varies with log format complexity
• Best used as part of broader security monitoring

🎯 Intended Use

Primary Applications

• Security Information and Event Management (SIEM)
• Log analysis automation
• Threat detection and classification
• Security monitoring workflows

Out of Scope

• Real-time critical security decisions without human oversight
• Standalone security solution
• Legal or compliance determinations

📝 Citation

If you use this model in your research or applications, please cite:

@misc{{cyberguard-ai-2024,
  author = {{{omasteam}}},
  title = {{CyberGuard AI: Fine-tuned Security Log Analyzer}},
  year = {{2024}},
  publisher = {{Hugging Face}},
  journal = {{Hugging Face Model Hub}},
  howpublished = {{\\url{{https://huggingface.co/{omasteam}/{cyberguard-ai-security-analyzer}}}}}
}}

🤝 Contributing

Contributions, suggestions, and feedback are welcome! Please open an issue or submit a pull request.

📄 License

This model is released under the Apache 2.0 License.

🙏 Acknowledgments

• Unsloth Team: For the efficient training framework
• Hugging Face: For the model hosting platform
• Meta AI: For the base Llama-2 model

---