/detecting-anomalies-in-industrial-control-systems

name: detecting-anomalies-in-industrial-control-systems
description: Deploys anomaly detection for OT/ICS environments using machine learning on OT network baselines, physics-based process models, and Modbus/DNP3/OPC UA traffic analysis to flag deviations, rogue devices, and mismatches against historian data. Use for continuous OT monitoring, baselining deterministic SCADA polling, or investigating alerts from Nozomi Guardian/Dragos needing deeper protocol analysis.
domain: cybersecurity
subdomain: ot-ics-security
tags:
- ot-security
- ics
- scada
- industrial-control
- iec62443
- anomaly-detection
- machine-learning
version: 1.0.0
author: mahipal
license: Apache-2.0
atlas_techniques:
- AML.T0043
- AML.T0018
nist_ai_rmf:
- MEASURE-2.7
- MEASURE-2.5
- MAP-5.1
nist_csf:
- PR.IR-01
- DE.CM-01
- ID.AM-05
- GV.OC-02
mitre_attack:
- T0836
- T0831
- T0832
- T0814
- T0801

Detecting Anomalies in Industrial Control Systems

When to Use

• When deploying continuous monitoring for OT environments that lack intrusion detection
• When building behavior-based detection to complement signature-based IDS in OT networks
• When establishing baselines for deterministic SCADA communications to detect deviations
• When integrating machine learning anomaly detection with OT security monitoring platforms
• When investigating alerts from Nozomi Guardian or Dragos Platform that require deeper analysis

**Do not use** for signature-based detection of known exploits (see detecting-attacks-on-scada-systems), for IT network anomaly detection without OT protocols, or as a replacement for process safety systems (SIS).

Prerequisites

• Passive network monitoring sensors on OT network SPAN/TAP ports
• Minimum 2-4 weeks of baseline traffic capture during normal operations
• Python 3.9+ with scikit-learn, numpy, pandas for ML model training
• Process historian access for physical process correlation data
• Understanding of normal operational patterns including shift changes, batch processes, and maintenance windows

Workflow

Step 1: Build Multi-Dimensional Baseline Model

Capture and model the deterministic behavior of ICS communications across multiple dimensions: timing, protocol behavior, and network topology.

#!/usr/bin/env python3
"""ICS Anomaly Detection System.

Builds multi-dimensional baselines from OT network traffic and
detects anomalies using statistical and machine learning methods.
Designed for deterministic SCADA communication patterns.
"""

import json
import sys
import time
import warnings
from collections import defaultdict
from datetime import datetime, timedelta
from dataclasses import dataclass, field

import numpy as np
import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.preprocessing import StandardScaler

warnings.filterwarnings("ignore")


@dataclass
class CommunicationProfile:
    """Profile for a single master-slave communication pair."""
    src_ip: str
    dst_ip: str
    protocol: str
    port: int
    avg_interval_ms: float = 0.0
    std_interval_ms: float = 0.0
    avg_payload_size: float = 0.0
    function_codes: dict = field(default_factory=dict)
    packets_per_minute: float = 0.0
    first_seen: str = ""
    last_seen: str = ""


class ICSAnomalyDetector:
    """Multi-dimensional anomaly detection for ICS environments."""

    def __init__(self):
        self.profiles = {}
        self.topology_baseline = set()
        self.timing_model = None
        self.isolation_forest = None
        self.scaler = StandardScaler()
        self.anomalies = []
        self.training_data = []

    def build_baseline_from_pcap(self, pcap_data):
        """Build baselines from parsed pcap data (list of flow records)."""
        print("[*] Building ICS communication baselines...")

        for flow in pcap_data:
            key = f"{flow['src']}->{flow['dst']}:{flow['port']}"

            if key not in self.profiles:
                self.profiles[key] = CommunicationProfile(
                    src_ip=flow["src"],
                    dst_ip=flow["dst"],
                    protocol=flow.get("protocol", "TCP"),
                    port=flow["port"],
                    first_seen=flow.get("timestamp", ""),
                )

            profile = self.profiles[key]
            profile.last_seen = flow.get("timestamp", "")

            # Track function codes for industrial protocols
            fc = flow.get("function_code")
            if fc is not None:
                profile.function_codes[fc] = profile.function_codes.get(fc, 0) + 1

            # Add to topology baseline
            self.topology_baseline.add((flow["src"], flow["dst"], flow["port"]))

        # Calculate interval statistics
        self._calculate_timing_stats(pcap_data)

        print(f"  Communication pairs: {len(self.profiles)}")
        print(f"  Topology entries: {len(self.topology_baseline)}")

    def _calculate_timing_stats(self, flows):
        """Calculate packet timing statistics per communication pair."""
        timestamps = defaultdict(list)
        for flow in flows:
            key = f"{flow['src']}->{flow['dst']}:{flow['port']}"
            ts = flow.get("timestamp_epoch")
            if ts:
                timestamps[key].append(ts)

        for key, ts_list in timestamps.items():
            if key in self.profiles and len(ts_list) > 1:
                ts_sorted = sorted(ts_list)
                intervals = [
                    (ts_sorted[i+1] - ts_sorted[i]) * 1000
                    for i in range(len(ts_sorted) - 1)
                ]
                self.profiles[key].avg_interval_ms = np.mean(intervals)
                self.profiles[key].std_interval_ms = np.std(intervals)
                duration_min = (ts_sorted[-1] - ts_sorted[0]) / 60
                if duration_min > 0:
                    self.profiles[key].packets_per_minute = len(ts_list) / duration_min

    def train_isolation_forest(self, features_df):
        """Train Isolation Forest mode