FL-DPCSA: Federated learning with differential privacy for cache side-channel attack detection in edge-based smart grids

Document Type

Article

Publication Title

E Prime Advances in Electrical Engineering Electronics and Energy

Abstract

Smart grid technology adoption at a fast pace has created new security vulnerabilities to cache side-channel attacks (CSAs) which threaten both user privacy and grid stability through edge computing devices. The current centralized detection methods need complete raw data collection, which leads to privacy risks and scalability limitations. The proposed PPFL framework provides distributed CSA detection across smart meters through a privacy-preserving federated learning approach that avoids data sharing. The solution uses differential privacy with ϵ= 1.0–5.0 to secure aggregation and a lightweight CNN-LSTM model, which results in 96.3% detection accuracy while maintaining data confidentiality. Real-world smart meter datasets from UK-DALE and REDD, together with simulation tests, show that the framework operates efficiently (2.1 s training latency/round), has minimal communication overhead (1.2 MB/round), and remains resistant to adversarial attacks (4.8% accuracy drop under evasion attempts). The proposed framework demonstrates linear scalability to 10,000+ devices while using 2.7 Wh energy per round, which makes it suitable for extensive smart grid implementations that follow GDPR and NIST cybersecurity standards.

DOI

10.1016/j.prime.2025.101057

Publication Date

9-1-2025

This document is currently not available here.

Share

COinS