FedST: Secure Federated Shapelet Transformation for Time Series Classification

TL;DR

FedST presents a secure federated shapelet transformation for time series classification, enabling cross-organization learning without data sharing. It defines a federated shapelet search kernel, analyzes security and efficiency bottlenecks, and proposes FL-tailored acceleration techniques including a secure dot-product protocol, sorting-based IG acceleration, and an F-stat trade-off, achieving substantial speedups while preserving accuracy and interpretability. Experimental results on 97 UCR datasets and synthetic data demonstrate FedST’s competitive accuracy relative to centralized baselines and significant efficiency gains, including up to thousands of times faster federated search with combined optimizations. The framework can be extended with differential privacy to protect outputs further, making FedST a practical, interpretable, privacy-preserving solution for time series classification in federated settings.

Abstract

This paper explores how to build a shapelet-based time series classification (TSC) model in the federated learning (FL) scenario, that is, using more data from multiple owners without actually sharing the data. We propose FedST, a novel federated TSC framework extended from a centralized shapelet transformation method. We recognize the federated shapelet search step as the kernel of FedST. Thus, we design a basic protocol for the FedST kernel that we prove to be secure and accurate. However, we identify that the basic protocol suffers from efficiency bottlenecks and the centralized acceleration techniques lose their efficacy due to the security issues. To speed up the federated protocol with security guarantee, we propose several optimizations tailored for the FL setting. Our theoretical analysis shows that the proposed methods are secure and more efficient. We conduct extensive experiments using both synthetic and real-world datasets. Empirical results show that our FedST solution is effective in terms of TSC accuracy, and the proposed optimizations can achieve three orders of magnitude of speedup.

Figures (20)

• Figure 1: Example of enabling federated learning to enrich the training time series data. A business who owns some training time series samples (blue) collaborates with the partners who have additional training samples (green) to jointly build the TSC model. They follow some secure protocols to avoid disclosing their private training data.
• Figure 2: Illustration of the shapelet-based features. A shapelet is a salient subsequence that represents a shape unique to certain classes. With a few shapelets of high distinguishing ability, each time series sample is transformed into a low-dimensional feature vector representing how similar (distant) the sample is to these shapelets. The classification is made and explained based on the few features rather than the abundant data points of the raw time series. In this example, the time series similar to shapelet 1 (orange) and distant to shapelet 2 (blue) are classified into class 1 and vice versa.
• Figure 3: An illustration of the FedST framework.
• Figure 4: Throughputs (#operations per second) of different MPC operations executed by three parties. Secure addition is much more efficient than the others because it is executed without communication catrina2010secure.
• Figure 5: Illustration of the Euclidean norm pruning and its information disclosure.

Theorems & Definitions (18)

• definition thmcounterdefinition: FL-enabled TSC problem
• definition thmcounterdefinition: Security
• definition thmcounterdefinition: Federated Shapelet Search, $\mathcal{F}_{FedSS}$
• definition thmcounterdefinition: Indicating Vector
• theorem 1
• proof : Proof Sketch
• theorem 2
• proof : Proof Sketch
• theorem 3
• theorem 4