Research on Data Poisoning Attack against Smart Grid Cyber-Physical System Based on Edge Computing.

Area of Science:

• Cybersecurity
• Machine Learning
• Smart Grid Technology

Background:

• Data poisoning attacks threaten machine learning models by contaminating training data.
• These attacks are particularly critical in smart grids, potentially causing fatal operational disruptions.
• Existing methods for data poisoning are computationally expensive, hindering deployment in resource-limited smart grid edge environments.

Purpose of the Study:

• To propose an efficient and cost-effective online data poisoning attack framework for smart grids.
• To mitigate the challenges of implementing complex optimization problems in edge computing environments.
• To enhance the efficiency and reduce the overhead of data poisoning attacks.

Main Methods:

• Developed an online data poisoning attack framework utilizing an online regression task model.
• Implemented a sample loss-based point selection strategy for targeted data pollution.
• Introduced a batch-polluting strategy with synchronous gradient ascent updates to reduce optimization iterations.

Main Results:

• The proposed framework effectively manipulates machine learning models by polluting incremental data streams.
• The sample loss strategy significantly improves attack targeting and efficiency compared to random selection.
• The batch-polluting strategy reduces inner optimization iterations and time overhead.

Conclusions:

• The novel online data poisoning framework demonstrates superior attack effectiveness and reduced overhead compared to baseline methods.
• The proposed strategies are crucial for enabling sophisticated attacks in resource-constrained smart grid edge settings.
• This research highlights the need for robust defenses against advanced data poisoning techniques in critical infrastructure.