Adversarial concept drift detection under poisoning attacks for robust data stream mining.

Area of Science:

• Machine Learning
• Data Science
• Cybersecurity

Background:

• Continuous learning from streaming data is challenging due to data volume and evolving patterns (concept drift).
• Existing concept drift detectors fail to distinguish between natural drift and adversarial poisoning attacks.
• Adversarial attacks can maliciously inject false data to degrade classification system performance.

Purpose of the Study:

• To propose a robust framework for concept drift detection resilient to adversarial and poisoning attacks.
• To introduce a taxonomy for adversarial concept drifts.
• To develop a novel measure for evaluating drift detectors under attack.

Main Methods:

• Developed a robust trainable concept drift detector based on an augmented Restricted Boltzmann Machine (RBM).
• Improved gradient computation and energy function within the RBM for enhanced performance.
• Introduced Relative Loss of Robustness (RLR) for performance evaluation under poisoning.

Main Results:

• The proposed framework effectively detects concept drift in adversarial scenarios.
• Experiments demonstrate high robustness and efficacy on both fully and sparsely labeled data streams.
• The RBM-based detector successfully differentiates between real and adversarial concept drift.

Conclusions:

• The novel framework provides a robust solution for concept drift detection in the presence of sophisticated attacks.
• The proposed methods and metrics advance the field of secure and reliable machine learning on data streams.
• This research is crucial for building trustworthy AI systems that operate in dynamic and potentially hostile environments.