Efficient lightweight privacy data anomaly detection solution with robust aggregation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Federated learning (FL) presents challenges in detecting anomalies in privacy-sensitive text due to model poisoning and efficiency demands.
• Existing methods often struggle to balance robustness, computational cost, and privacy preservation in FL settings.

Purpose of the Study:

• To develop an integrated framework for robust and efficient anomaly detection in privacy-sensitive text within federated learning environments.
• To enhance the resilience of FL against model poisoning attacks while reducing inference costs.

Main Methods:

• An early-exit RoBERTa classifier (E2-RoBERTa) with multi-stage exits and a spatio-temporal convolutional-LSTM fusion module was developed.
• A robust federated layered aggregation strategy (RFLA) was proposed, incorporating dimensionality reduction, density clustering, and Mahalanobis-based weighting for server-side resilience.

Main Results:

• E2-RoBERTa achieved high detection accuracy, with experiments showing [Formula: see text] on SMS data.
• RFLA demonstrated superior accuracy and robustness compared to Krum, Trimmed Mean, and Median under significant malicious client percentages (20-50%).
• The early-exit mechanism reduced average inference time by approximately 17%.

Conclusions:

• The integrated framework effectively balances privacy preservation, robustness against model poisoning, and computational efficiency.
• The proposed E2-RoBERTa and RFLA offer a practical solution for privacy-text anomaly detection in federated settings.
• This approach supports the deployment of secure and efficient anomaly detection systems in sensitive data environments.