CloudAnoBench

Paper: Towards Generalizable Context-aware Anomaly Detection: A Large-scale Benchmark in Cloud Environments

Project Page: https://jayzou3773.github.io/cloudanobench-agent/

CloudAnoBench is a large-scale benchmark for context-aware anomaly detection in cloud environments, jointly incorporating both metrics and logs to more faithfully reflect real-world conditions. It consists of 1,252 labeled cases spanning 28 anomalous scenarios and 16 deceptive normal scenarios (approximately 200K lines), with explicit annotations for anomaly type and scenario. By including deceptive normal cases where anomalous-looking metric patterns are explained by benign log events, the benchmark introduces higher ambiguity and difficulty compared to prior datasets, thereby providing a rigorous testbed for evaluating both anomaly detection and scenario identification.

📁 Dataset Structure

The dataset is organized as follows:

cloudanobench/
├── anom_dataset/          # Anomaly scenarios (without Malicious)
│   ├── scenario_1/        # CPU Hog Process
│   ├── scenario_2/        # Memory Leak
│   └── ...
├── mali_dataset/          # Malicious Events scenarios
│   ├── scenario_1/        # Cryptomining Malware
│   ├── scenario_2/        # DDoS Botnet Agent
│   └── ...
└── norm_dataset/          # Deceptive Normal scnarios
    ├── scenario_1/        # Nightly Full Backup
    ├── scenario_2/        # Log Rotation & Compression
    └── ...

Each scenario directory contains:

• *.csv files: Structured performance metrics and log data
• *.log files: Raw log files

🚨 Anomaly scenarios (anom_dataset)

The anomaly scenarios contains 17 different types of system anomaly scenarios, organized by the following categories:

Resource Exhaustion & Bottlenecks

• Scenario 1: CPU Hog Process - Resource exhaustion caused by CPU-intensive processes
• Scenario 2: Memory Leak - Memory leakage issues
• Scenario 3: Disk I/O Bottleneck - Disk I/O performance bottlenecks
• Scenario 4: Noisy Neighbor - Noisy neighbor effects in shared environments
• Scenario 5: Handle/Thread Exhaustion - Handle or thread pool exhaustion

Network Anomalies

• Scenario 6: High Network Latency / Packet Loss - High network latency and packet loss
• Scenario 7: Bandwidth Saturation - Network bandwidth saturation
• Scenario 8: DNS Resolution Failure - DNS resolution failures
• Scenario 9: Firewall/Security Group Misconfiguration - Firewall or security group misconfigurations

Software & Application Anomalies

• Scenario 10: Application Crash Loop - Application crash-restart loops
• Scenario 11: Deadlock / Livelock - Deadlock and livelock situations
• Scenario 12: External Dependency Failure - External service dependency failures
• Scenario 13: Application Misconfiguration - Application configuration errors
• Scenario 14: Garbage Collection (GC) Storm - Excessive garbage collection activity

Other Complex & Subtle Anomalies

• Scenario 15: Time Skew (Clock Drift) - System clock drift and time synchronization issues
• Scenario 16: Kernel / Driver Bug - Kernel or device driver bugs
• Scenario 17: (Additional anomaly scenario)

🦠 Malicious Events scenarios Dataset (mali_dataset)

The malicious activity dataset contains 12 different types of malicious event scenarios:

Malicious Events

• Scenario 1: Cryptomining Malware - Cryptocurrency mining malware activities
• Scenario 2: DDoS Botnet Agent - DDoS botnet agent operations
• Scenario 3: Ransomware - Ransomware infection and encryption activities
• Scenario 4: Data Exfiltration - Unauthorized data theft and transfer
• Scenario 5: Spam Bot - Spam email distribution bot activities
• Scenario 6: Rootkit / Backdoor - Rootkit installation and backdoor access
• Scenario 7: Brute-force / Credential Stuffing Attack - Password brute-force and credential stuffing attacks
• Scenario 8: Web Shell Beaconing & Command Execution - Web shell communication and remote command execution
• Scenario 9: Password Cracking - Password cracking and hash breaking activities
• Scenario 10: Log Deletion / Tampering - System log deletion and tampering
• Scenario 11: Reverse Shell - Reverse shell establishment and communication
• Scenario 12: Application-Layer DDoS Attack - Application-layer DDoS attacks

✅ Deceptive Normal scnarios Dataset (norm_dataset)

The normal behavior dataset contains 16 different normal business operation scenarios:

Normal Business Operations

• Scenario 1: Nightly Full Backup - Scheduled full system backup operations
• Scenario 2: Log Rotation & Compression - Automated log file rotation and compression
• Scenario 3: Periodic Data Aggregation - Regular data aggregation and summarization tasks
• Scenario 4: Filesystem Cleanup - Routine filesystem maintenance and cleanup
• Scenario 5: SSL Certificate Renewal - Automated SSL/TLS certificate renewal processes
• Scenario 6: Blue-Green Deployment - Blue-green application deployment procedures
• Scenario 7: Live Streaming Event Start - Live streaming service initialization
• Scenario 8: Search Engine Aggressive Crawl - Intensive web crawling operations
• Scenario 9: End-of-Month Financial Closing - Month-end financial processing workflows
• Scenario 10: On-Demand Full Report Generation - Large-scale report generation tasks
• Scenario 11: ETL Pipeline Execution - Extract, Transform, Load data pipeline operations
• Scenario 12: On-Demand Video Transcoding - Video transcoding and processing tasks
• Scenario 13: Database Compaction/Vacuum - Database maintenance and optimization
• Scenario 14: Cache Eviction Storm - Cache invalidation and eviction processes
• Scenario 15: Connection Pool Scaling - Dynamic connection pool scaling operations
• Scenario 16: File Integrity Monitoring Scan - System-wide file integrity verification

🔧 Usage

Data Loading Example

import pandas as pd
import os

# Load data for a specific scenario
def load_scenario_data(dataset_type, scenario_id):
    """
    Load data for a specific scenario
    
    Args:
        dataset_type: 'anom', 'mali', or 'norm'
        scenario_id: Scenario number (1-17 for anom, 1-12 for mali, 1-16 for norm)
    """
    base_path = f"{dataset_type}_dataset/scenario_{scenario_id}"
    
    csv_files = []
    log_files = []
    
    for file in os.listdir(base_path):
        if file.endswith('.csv'):
            csv_files.append(pd.read_csv(os.path.join(base_path, file)))
        elif file.endswith('.log'):
            log_files.append(os.path.join(base_path, file))
    
    return csv_files, log_files

# Example: Load CPU Hog Process anomaly data
csv_data, log_files = load_scenario_data('anom', 1)

Batch Data Processing

# Batch load all anomaly scenarios
anomaly_data = {}
for i in range(1, 18):  # scenarios 1-17
    anomaly_data[f'scenario_{i}'] = load_scenario_data('anom', i)

# Batch load all malicious activity scenarios  
malicious_data = {}
for i in range(1, 13):  # scenarios 1-12
    malicious_data[f'scenario_{i}'] = load_scenario_data('mali', i)

# Batch load all normal behavior scenarios
normal_data = {}
for i in range(1, 17):  # scenarios 1-16
    normal_data[f'scenario_{i}'] = load_scenario_data('norm', i)

📋 File Naming Convention

• Anomaly data: anom_{scenario_id}_{case_number}.{csv|log}
• Malicious data: mali_{scenario_id}_{case_number}.{csv|log}
• Normal data: norm_{scenario_id}_{case_number}.{csv|log}

Where:

• scenario_id: Scenario number
• case_number: Case number within the scenario
• Extension: .csv for structured data, .log for raw logs

@misc{zou2025generalizablecontextawareanomalydetection,
  title={Towards Generalizable Context-aware Anomaly Detection: A Large-scale Benchmark in Cloud Environments}, 
  author={Xinkai Zou and Xuan Jiang and Ruikai Huang and Haoze He and Parv Kapoor and Hongrui Wu and Yibo Wang and Jian Sha and Xiongbo Shi and Zixun Huang and Jinhua Zhao},
  year={2025},
  eprint={2508.01844},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2508.01844}, 
}