TEE-based Selective Testing of Local Workers in Federated Learning Systems

Wensheng Zhang, Trent Muhr

Abstract: This paper considers a federated learning system composed of a central coordinating server and multiple distributed local workers, all having access to trusted execution environments (TEEs). In order to ensure that the untrusted workers correctly perform local learning, we propose a new TEE-based approach that also combines techniques from applied cryptography, smart contract and game theory. Theoretical analysis and implementation-based evaluations show that, the proposed approach is secure, efficient and practical.

Introduction

Federated Learning (FL) enables collaborative model training across distributed devices while preserving data privacy. However, the decentralized nature of FL introduces a significant trust gap. "Lazy" workers may skip computationally expensive training steps to conserve resources, while malicious actors might submit poisoned updates. Detecting such behavior is challenging because the central server lacks visibility into local data and training processes.

Traditional solutions often propose executing the entire training workload within a Trusted Execution Environment (TEE), such as Intel SGX. However, TEEs suffer from significant performance overheads and strict memory limitations (e.g., the 128MB Enclave Page Cache in older SGX versions), making them impractical for modern deep learning workloads. This paper introduces a "selective testing" framework that leverages TEEs for minimal auditing rather than full execution.

System Architecture

The proposed system integrates three primary components: an Aggregation Server, a Blockchain-based smart contract layer, and multiple Local Workers equipped with TEEs.

Figure 1: Overview of the TEE-based Federated Learning architecture, showing the interaction between the aggregation server, blockchain, and local worker TEEs.

As illustrated in Figure 1, the architecture utilizes a Trusted Local Monitor (TLM) within each worker's TEE. The TLM does not perform the training itself but acts as an on-device auditor. The blockchain layer manages financial deposits, providing a mechanism for game-theoretic deterrence against dishonest behavior.

The Selective Testing Framework

The core of the framework is TEE-based Selective Testing, which rests on three pillars:

  1. Selective Auditing: Instead of verifying the entire training process, the TLM randomly selects a minute fraction of operations (e.g., specific layer passes) for verification.
  2. Cryptographic Commitments: To prevent workers from anticipating which operations will be audited, they must first commit to their results using Merkle Hash Trees. Only after this commitment does the TLM reveal the audit targets.
  3. Game-Theoretic Deterrence: Workers are modeled as economically rational agents. By requiring a financial deposit that is forfeited upon detection of cheating, the system ensures that even a low probability of auditing (e.g., testing only two operations) is sufficient to deter dishonesty.

Experimental Evaluation

The researchers benchmarked the system using an Intel Core i5-8400 CPU with SGX support, focusing on the latency of convolutional and fully-connected layers across various input scales.

Figure 2: Performance comparison (latency in µs) across different input sizes, highlighting the efficiency of the selective test compared to full SGX execution.

As shown in Figure 2, the "selective test" performed within the TEE is remarkably efficient. For a 256×256 input, the selective audit takes only 360µs, whereas running the full forward pass in SGX ("SGX fwd") requires 76,868µs. While the total overhead of the new scheme ("new fwd") is higher than the unsecured version ("original fwd") due to Merkle tree construction, the increase is manageable, ranging from 1.14x to 1.55x. Crucially, the relative cost of the TEE audit decreases as layer complexity increases, demonstrating excellent scalability.

Conclusion

This research demonstrates that full TEE execution is not a prerequisite for security in Federated Learning. By combining lightweight TEE-based auditing with cryptographic commitments and economic incentives, the proposed framework achieves robust security with minimal performance penalties. Future work will focus on extending these verification methods to more complex architectures like Transformers and optimizing the cryptographic commitment process to further reduce non-TEE overhead.

Summary updated

Abstract

This paper considers a federated learning system composed of a central coordinating server and multiple distributed local workers, all having access to trusted execution environments (TEEs). In order to ensure that the untrusted workers correctly perform local learning, we propose a new TEE-based approach that also combines techniques from applied cryptography, smart contract and game theory. Theoretical analysis and implementation-based evaluations show that, the proposed approach is secure, efficient and practical.
BibTeX
@misc{zhang2021tee,
  author = {Wensheng Zhang and Trent Muhr},
  title = {{TEE-based Selective Testing of Local Workers in Federated Learning Systems}},
  year = {2021},
  eprint = {2111.02662},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.02662},
}
Figure