Identifying alternately poisoning attacks in federated learning online using trajectory anomaly detection method.

Area of Science:

• Machine Learning
• Cybersecurity
• Data Science

Background:

• Federated learning (FL) faces implicit poisoning attacks where malicious nodes subtly alter gradients.
• Existing detection methods, relying on temporal analysis or offline processing, are often ineffective or delayed.
• Covert attacks can bypass common online detection techniques like cosine similarity and clustering.

Purpose of the Study:

• To investigate the challenges of detecting implicit poisoning in federated learning.
• To develop a novel, privacy-preserving algorithm for identifying malicious clients in FL systems.
• To enhance the security and integrity of federated learning models against subtle manipulation.

Main Methods:

• Recalculating gradient updates to reveal distinct malicious client characteristics.
• Designing a privacy-preserving detection algorithm based on trajectory anomaly detection.
• Utilizing singular values of matrices as features and an improved Isolation Forest for anomaly detection.

Main Results:

• The proposed method achieved 94.3% detection accuracy on benchmark datasets (MNIST, FashionMNIST, CIFAR-10).
• The false positive rate was maintained below 1.2%, demonstrating high precision.
• The algorithm effectively identifies malicious behavior indicative of implicit model poisoning attacks.

Conclusions:

• Temporal analysis alone is insufficient for detecting sophisticated implicit poisoning attacks in FL.
• Recalculating gradient updates offers a viable approach to uncover malicious activities.
• The developed privacy-preserving trajectory anomaly detection method is accurate and effective for securing federated learning.