TL;DR
This paper introduces DeeBBAA, a novel black-box adversarial attack strategy that targets the AC state estimation in cyber-physical power systems. By using a substitute neural network trained on limited historical data, DeeBBAA can generate stealthy false data injection attacks that effectively evade a wide range of conventional and advanced detection mechanisms, posing a significant threat to grid security.
What Problem Does This Paper Address?
Modern power systems are increasingly becoming "smart" and cyber-physical, integrating advanced communication and automation technologies. While this enhances efficiency and reliability, it also introduces vulnerabilities to cyberattacks. A critical operation in power systems is State Estimation (SE), which uses sensor measurements to determine the real-time operating status of the grid. False Data Injection (FDI) attacks aim to compromise the integrity of these measurements, leading to incorrect state estimates, potential load shedding, economic losses, and even blackouts.
Existing FDI attack strategies often require extensive knowledge of the power system's topology, parameters, and even the detection mechanisms in place, making them impractical. Furthermore, many attacks are designed against simplified DC power flow models or are tested against a limited set of detection algorithms. This paper addresses the need for a more realistic and challenging attack that operates in a black-box setting (without full knowledge of the target system) and can bypass a broad spectrum of modern detection methods, including those based on physics and machine learning.
Methods
The core of DeeBBAA is a two-stage process:
- Substitute Model Training (Offline): The attacker first identifies an "attack region" within the power system. Using historical power measurements (bus injections and line flows) from this region, a Neural State Estimator (NSE) is trained. This NSE acts as a substitute model that mimics the behavior of the unknown AC state estimation module for the identified region. This stage requires only partial historical data and no direct interaction with the target system.
- Adversarial Attack Generation (Online): Once the NSE is trained, the attacker collects real-time power measurements from the attack region. A convex optimization problem is formulated to find an adversarial perturbation that, when added to these measurements, maximizes the deviation in the NSE's output while staying within a defined perturbation budget ($\epsilon$). This optimization problem is solved using a novel convex relaxation technique involving Taylor series approximation and Semi-Definite Programming (SDP). The resulting perturbation is then injected into the real-time measurements, creating stealthy false data.
A key aspect is the use of a substitute model (NSE) that is trained on a different dataset than the one used for training the defense baselines. This prevents data spillage and ensures a more robust evaluation of the attack's effectiveness. The attack also allows for flexible control over its magnitude and spread by adjusting parameters like $\epsilon$ and the measurement selection vector.
Datasets
- Dataset A: Generated from the Australian Energy Market Operator (AEMO) load profiles (2019) and interpolated to a 5-minute interval. This dataset, representing normal operating conditions, was used to train the Neural State Estimator (NSE).
- Dataset B: Generated from the New York Independent System Operator (NYISO) load profiles (2021) with a 5-minute interval. This dataset was used for:
- Training the defense baselines.
- Generating online DeeBBAA attacks.
- Dataset Batt: A subset of Dataset B, specifically generated to contain standard Stealthy False Data Injection (SFDI) attacks. This dataset was used to train the defense algorithms to recognize known attack patterns.
The paper uses the standard IEEE 39-bus and 118-bus test systems for simulations.
Key Insights and Ideas
- Black-Box Transferability: The core idea is to train a substitute model (NSE) on historical data and then use it to generate attacks. The effectiveness of these attacks against the actual target system (even with a different architecture for the NSE) demonstrates the transferability of adversarial examples in this domain.
- Convex Relaxation for Attack Generation: The paper introduces a sophisticated method to convert a non-convex optimization problem for finding adversarial perturbations into a solvable convex problem (SDP). This is crucial for efficiently generating attacks in real-time.
- Data Independence for Defense Training: Training the NSE on a completely separate dataset from the defense baselines is a critical design choice. It ensures that the attack is not implicitly "known" to the defenses and provides a more realistic worst-case scenario for evaluation.
- Versatility of Attack Regions: The ability to define both "localized" and "delocalized" attack regions provides flexibility for the attacker, allowing them to adapt to different reconnaissance capabilities and target specific vulnerabilities.
Key Results
- High Evasion Rates: DeeBBAA attacks demonstrated a very high probability of bypassing a wide range of defense mechanisms.
- Conventional BDD (LNRT and $\chi^2$ test): Bypassed with at least 90% probability in most cases, and up to 40% detection only under specific high-epsilon conditions.
- Statistical consistency checks (KLD, Transformed KLD, ANCUSUM-C): Bypassed with almost 100% probability.
- Learning-based defenses (MLP, LSTM, Bi-LSTM, TCN, ARMA-GNN, E3LM, DAGMM): Bypassed with average probabilities ranging from 70% to over 90%, depending on the defense and attack configuration. Notably, MLP was completely ineffective.
- Significant State Deviations: The attacks successfully induced substantial deviations in estimated voltage magnitudes (up to 0.1 pu median) and angles (up to 2 degrees median) in the IEEE 39 bus system, and even higher deviations in the IEEE 118 bus system, under localized attack settings. Delocalized attacks led to even greater state deviations.
- Scalability and Efficiency: The proposed method is computationally efficient due to the convex relaxation, and the attack can be scaled by adjusting parameters like $\epsilon$ and the number of compromised measurements.
Limitations and Future Work
- Data Requirement: While DeeBBAA requires only partial historical data, it still relies on the availability of such data for reconnaissance.
- NSE Architecture: The paper notes that the MLP-based NSE is a simplified model. While effective for generating attacks, its architecture differs from typical complex state estimation models, which might influence transferability in some scenarios.
- Future Work: The paper highlights the grave vulnerability of current detection mechanisms and emphasizes the need for a combination of data security, communication channel security, and stronger attack detection mechanisms. Future work could explore more sophisticated NSE architectures, adaptive attack strategies that evolve over time, and defenses that are inherently more robust to such black-box attacks.
Summary updated