Developing a Strong CPS Defender: An Evolutionary Approach

1. Problem Statement

Cyber-Physical Systems (CPSs) are critical infrastructure components (e.g., water treatment, power grids) that face severe security threats. Traditional anomaly detection methods suffer from several limitations:

• Data Scarcity & Bias: Training datasets are often limited to benign operating conditions or static, expert-crafted attack scenarios, failing to generalize to unseen or subtle attack strategies.
• Static Defenses: Most detectors are trained offline (batch learning) and cannot adapt to evolving threats or sensor drift without expensive retraining from scratch.
• Resource Constraints: Collecting diverse attack data requires significant domain expertise and computational resources, which defenders often lack.
• Redundancy: Automated fuzzing techniques often generate redundant or noisy data, focusing solely on attack success rather than generating high-quality, diverse training data for defenders.

The core challenge is to build a robust, adaptive defender that can learn from limited data and continuously improve against evolving attack strategies without requiring massive historical datasets.

2. Methodology: Evo-Defender Framework

The authors propose Evo-Defender, an evolutionary framework that simulates a dynamic interaction between a "Spear" (smart attacker) and a "Shield" (adaptive defender). The system operates iteratively to strengthen defenses.

A. The Spear (Smart Attacker)

The goal of the Spear is not just to find any attack, but to generate diverse, non-redundant, and impactful attack vectors to expand the attack surface coverage.

• Predictive Modeling: A machine learning model (LSTM) is trained on historical CPS data to predict system states based on sensor/actuator inputs. This guides the attack generation.
• Guided Fuzzing: The Spear uses a coverage-guided fitness function to select attacks. This function evaluates:
  1. Diversity: How different the new attack is from historical attacks (using embedding distances).
  2. Impact: How far the attack deviates the system from normal operational constraints (e.g., pushing pressure or temperature to safety limits).
• Search Strategy: It employs random search and Genetic Algorithms (GA) to explore the space of sensor/actuator manipulations, selecting candidates via Roulette Wheel Selection based on fitness scores.

B. The Shield (Incremental Defender)

The Shield is an anomaly detection model that evolves continuously using incremental learning. It avoids retraining from scratch by processing misclassified data in real-time.

• Data Collection: The system identifies "misclassified data points" (False Negatives where attacks were missed, or False Positives where normal operation was flagged).
• Three-Pronged Incremental Learning Strategy:
  1. Exemplar Replay (Catastrophic Forgetting Mitigation): To prevent the model from forgetting old patterns when learning new ones, the system maintains a small set of "exemplars" (representative samples from previous rounds) selected based on feature similarity. These are replayed during training.
  2. Class Balance Loss (Data Imbalance): Since attack data is scarce compared to normal data, a penalty term is added to the Binary Cross-Entropy loss to balance the learning weight between normal and abnormal classes.
  3. Continual Backpropagation (Plasticity Loss Mitigation): To prevent the model from becoming rigid over time, the system dynamically identifies "inactive" neurons (those with low activation/utility) and reinitializes their weights. This injects controlled randomness to maintain the model's ability to adapt to new data.

3. Key Contributions

1. Automated Attack Generation: Development of a smart attacker that generates over 600 diverse attack scenarios across two real-world CPS platforms, ensuring comprehensive coverage of the attack surface.
2. Continual Learning Defender: The first online scheme for incrementally updating deep learning-based anomaly detectors in CPS. It addresses the "data scarcity" and "resource constraint" paradox by learning from misclassified samples without full retraining.
3. Robust Detection Capability: Demonstration of a system that outperforms state-of-the-art baselines in both accuracy and generalization to unseen attacks, while significantly reducing data consumption.
4. Open Data: Release of the generated attack datasets and results for reproducibility.

4. Experimental Results

The framework was evaluated on two testbeds:

• Tennessee Eastman (TE): A simulated chemical process.
• Robotic Arm Assembly Workstation (RAAW): A real-world robotic platform with a PLC.

Key Findings:

• Performance Gain: Evo-Defender achieved up to 2.7× higher performance than state-of-the-art baselines (CNN, GRU, LSTM) on unseen attack scenarios.
  • TE Platform: 92.03% accuracy (vs. ~68% for baselines) and a false positive rate reduced to one-third of the baseline.
  • RAAW Platform: 89.59% accuracy (vs. ~60% for baselines) and a detection rate nearly tripled compared to the best baseline.
• Data Efficiency: The approach required 69.6% less training data to reach the same performance level as traditional methods.
• Generalization: In end-to-end tests, Evo-Defender correctly identified 89/103 unseen attacks on TE and 38/44 on RAAW, whereas baselines struggled with high false positives or missed attacks.
• Ablation Study: Confirmed that all three modules (Exemplars, Class Balance, Continual Backpropagation) are essential. Removing any module led to catastrophic forgetting, instability, or slower convergence.

5. Significance

This work addresses a critical gap in CPS security by shifting from static, one-off training to a dynamic, evolutionary defense mechanism.

• Practicality: It offers a solution for legacy systems where collecting massive labeled attack datasets is impossible.
• Adaptability: By simulating an active attacker, the system proactively discovers vulnerabilities and hardens the defense before real-world attacks occur.
• Scalability: The incremental learning approach allows defenders to scale to complex, data-scarce environments without prohibitive computational costs.

In summary, Evo-Defender demonstrates that an adversarial, evolutionary approach can significantly enhance the resilience of CPSs, providing a practical pathway for securing critical infrastructure against sophisticated, evolving threats.