Proposing Chaos Engineering Framework for Threat Detection Pipelines Using Deep Learning Based Hybrid CNN-LSTM

Abstract Chaos engineering is an emerging discipline that proactively tests system resilience by deliberately injecting controlled failures and anomalies into operational environments. This study addresses the application of chaos engineering principles to cybersecurity threat detection pipelines, specifically targeting Security Information and Event Management (SIEM) systems and Intrusion Detection Systems (IDS). The proposed framework integrates a Hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) model with chaos-augmented training, evaluated on the CICIDS2017 dataset comprising over 2.8 million network flow records across 15 attack categories. Our methodology incorporates chaos injection techniques including network partition simulation, log corruption, traffic replay, and resource exhaustion to evaluate pipeline robustness under adversarial conditions. The solution strategies focus on improving detection accuracy and reducing false positive rates while maintaining pipeline resilience under chaotic conditions. This study is simulated in Python using approximately 2,830,743 network flow records with 78 features from the CICIDS2017 benchmark dataset. Our goal is to compare the performance of our proposed Hybrid CNN-LSTM model against traditional ML-based approaches, highlighting advantages, limitations, and real-world applicability. The results demonstrated that chaos-augmented deep learning pipelines significantly enhance threat detection, achieving 92% overall accuracy and outperforming conventional ML approaches by 4%. Keywords:CHAOS ENGINEERING, THREAT DETECTION, CNN-LSTM, CICIDS2017, INTRUSION DETECTION SYSTEM, CYBERSECURITY, ANOMALY DETECTION, SIEM, DEEP LEARNING, NETWORK SECURITY.