train.py

import os
os.environ["CUDA_VISIBLE_DEVICES"]="3"
import torch
from datasets import Dataset
from unsloth import FastLanguageModel
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler
from transformers import TrainingArguments
from trl import SFTTrainer
import logging
from typing import List, Dict
import json

def setup_logging():
    logging.basicConfig(
        level=logging.INFO,
        format='%(asctime)s - %(levelname)s - %(message)s'
    )
    return logging.getLogger(__name__)
def sample_balanced_dataset(df, max_samples_per_class=1000):
    """Sample a balanced subset of the data"""
    sampled_dfs = []
    
    # Handle 'benign' class separately
    benign_df = df[df['Attack'].str.lower() == 'benign']
    attack_df = df[df['Attack'].str.lower() != 'benign']
    
    # Sample benign data
    if len(benign_df) > max_samples_per_class:
        benign_sampled = benign_df.sample(n=max_samples_per_class, random_state=42)
        sampled_dfs.append(benign_sampled)
    else:
        sampled_dfs.append(benign_df)
    
    # Sample each attack type
    for attack_type in attack_df['Attack'].unique():
        attack_type_df = attack_df[attack_df['Attack'] == attack_type]
        if len(attack_type_df) > max_samples_per_class:
            sampled = attack_type_df.sample(n=max_samples_per_class, random_state=42)
            sampled_dfs.append(sampled)
        else:
            sampled_dfs.append(attack_type_df)
    
    return pd.concat(sampled_dfs, ignore_index=True)

class NetworkFlowDataProcessor:
    def __init__(self):
        self.logger = setup_logging()
        self.scaler = MinMaxScaler()
        self.numerical_features = [
            'L4_SRC_PORT', 'L4_DST_PORT', 'PROTOCOL', 'L7_PROTO',
            'IN_BYTES', 'OUT_BYTES', 'IN_PKTS', 'OUT_PKTS',
            'TCP_FLAGS', 'FLOW_DURATION_MILLISECONDS'
        ]
        self.categorical_features = ['IPV4_SRC_ADDR', 'IPV4_DST_ADDR']
        
    def process_ip_address(self, ip: str) -> str:
        """Convert IP address to a more descriptive format"""
        parts = ip.split('.')
        if parts[0] == '192' and parts[1] == '168':
            return f"internal_network_{parts[2]}_{parts[3]}"
        return f"external_network_{ip}"

    def format_flow_data(self, row: pd.Series) -> str:
        """Format network flow data into a descriptive text"""
        return f"""Network Flow Description:
Source: {self.process_ip_address(row['IPV4_SRC_ADDR'])} (Port: {row['L4_SRC_PORT']})
Destination: {self.process_ip_address(row['IPV4_DST_ADDR'])} (Port: {row['L4_DST_PORT']})
Protocol Information:
- Protocol ID: {row['PROTOCOL']}
- Layer 7 Protocol: {row['L7_PROTO']}
- TCP Flags: {row['TCP_FLAGS']}
Traffic Metrics:
- Bytes: {row['IN_BYTES']} inbound, {row['OUT_BYTES']} outbound
- Packets: {row['IN_PKTS']} inbound, {row['OUT_PKTS']} outbound
- Duration: {row['FLOW_DURATION_MILLISECONDS']} milliseconds"""

    def get_attack_description(self, attack_type: str) -> str:
        """Get detailed description of attack type"""
        descriptions = {
            "benign": "This is normal network traffic with no malicious intent.",
            "ddos": "A Distributed Denial of Service attack attempting to overwhelm network resources.",
            "dos": "A Denial of Service attack targeting system availability.",
            "injection": "An attack attempting to inject malicious code or commands.",
            "scanning": "Network scanning activity to discover vulnerabilities.",
            "backdoor": "Malicious activity indicating backdoor access attempts.",
            "mitm": "Man-in-the-Middle attack intercepting network communications.",
            "password": "Password-based attack attempting unauthorized access.",
            "ransomware": "Ransomware-related network activity.",
            "xss": "Cross-Site Scripting attack targeting web applications."
        }
        return descriptions.get(attack_type.lower(), "Unknown attack type")

    def prepare_training_text(self, row: pd.Series) -> str:
        """Prepare single training example in LLaMA-3 chat format"""
        flow_text = self.format_flow_data(row)
        attack_type = row['Attack'].lower() if 'Attack' in row else 'benign'
        attack_desc = self.get_attack_description(attack_type)
        
        return f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
Analyze this network flow for potential security threats:

{flow_text}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
This network flow is classified as {attack_type}. {attack_desc}

Key indicators from the flow data:
- Traffic volume: {row['IN_BYTES'] + row['OUT_BYTES']} total bytes
- Flow duration: {row['FLOW_DURATION_MILLISECONDS']} ms
- Protocol behavior: {row['TCP_FLAGS']} TCP flags<|eot_id|>"""

def load_and_process_data(train_path: str, processor: NetworkFlowDataProcessor, max_samples_per_class=50000):
    """Load and process the training data"""
    logger = setup_logging()
    logger.info(f"Loading data from {train_path}")
    
    df = pd.read_csv(train_path)
    df = sample_balanced_dataset(df, max_samples_per_class)
    logger.info(f"Sampled dataset size: {len(df)}")
    
    # Create dataset with text field
    texts = [processor.prepare_training_text(row) for _, row in df.iterrows()]
    dataset = Dataset.from_pandas(pd.DataFrame({'text': texts}))
    
    return dataset

def main():
    logger = setup_logging()
    
    # Initialize data processor
    processor = NetworkFlowDataProcessor()
    
    # Load and process data
    train_dataset = load_and_process_data("data/train.csv", processor, max_samples_per_class=50000)
    
    # Model initialization
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name="unsloth/llama-3-8b-Instruct-bnb-4bit",
        max_seq_length=2048,
        load_in_4bit=True,
    )
    
    # Configure tokenizer for LLaMA-3
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"
    
    # Add LoRA adapters
    model = FastLanguageModel.get_peft_model(
        model,
        r=16,
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj",
            "gate_proj", "up_proj", "down_proj",
        ],
        lora_alpha=16,
        lora_dropout=0,
        bias="none",
        use_gradient_checkpointing="unsloth",
        random_state=3407,
    )
    
    # Configure training arguments
    training_args = TrainingArguments(
        output_dir="cybersec_model_output",
        num_train_epochs=3,
        per_device_train_batch_size=64,
        gradient_accumulation_steps=4,
        learning_rate=2e-4,
        bf16=True,  # Use bfloat16 for A100
        logging_steps=10,
        save_strategy="epoch",
        optim="adamw_8bit",
        lr_scheduler_type="cosine",
    )
    
    # Initialize trainer
    trainer = SFTTrainer(
        model=model,
        tokenizer=tokenizer,
        train_dataset=train_dataset,
        dataset_text_field="text",  # Changed from conversations to text
        max_seq_length=2048,
        args=training_args,
    )
    
    # Train the model
    logger.info("Starting training...")
    trainer.train()
    
    # Save the model
    logger.info("Saving model...")
    model.save_pretrained("cybersec_model")
    tokenizer.save_pretrained("cybersec_model")

if __name__ == "__main__":
    main()