[Paper Review] SecFL: Confidential Federated Learning using TEEs

Do Le Quoc, Christof Fetzer|arXiv (Cornell University)|Oct 3, 2021

Privacy-Preserving Technologies in Data4 citations

TL;DR

SecFL is a confidential federated learning framework that leverages Trusted Execution Environments (TEEs), such as Intel SGX, to ensure end-to-end confidentiality and integrity of model training. By executing local and global training within isolated enclaves and using remote attestation to detect malicious clients, SecFL maintains model accuracy while protecting against data leakage, model stealing, and Byzantine attacks.

ABSTRACT

Federated Learning (FL) is an emerging machine learning paradigm that enables multiple clients to jointly train a model to take benefits from diverse datasets from the clients without sharing their local training datasets. FL helps reduce data privacy risks. Unfortunately, FL still exist several issues regarding privacy and security. First, it is possible to leak sensitive information from the shared training parameters. Second, malicious clients can collude with each other to steal data, models from regular clients or corrupt the global training model. To tackle these challenges, we propose SecFL - a confidential federated learning framework that leverages Trusted Execution Environments (TEEs). SecFL performs the global and local training inside TEE enclaves to ensure the confidentiality and integrity of the computations against powerful adversaries with privileged access. SecFL provides a transparent remote attestation mechanism, relying on the remote attestation provided by TEEs, to allow clients to attest the global training computation as well as the local training computation of each other. Thus, all malicious clients can be detected using the remote attestation mechanisms.

Motivation & Objective

• Address privacy and security vulnerabilities in federated learning, including model and data leakage from shared parameters.
• Mitigate risks from malicious clients that collude to steal data, manipulate models, or corrupt the global model.
• Provide end-to-end confidentiality and integrity for training data, code, and model parameters during federated training.
• Enable trust among untrusted clients through transparent remote attestation of enclave execution.
• Maintain high model utility and accuracy while defending against powerful adversaries with privileged access.

Proposed method

• Executes local and global training computations inside Intel SGX enclaves to isolate sensitive data and code from untrusted environments.
• Uses SCONE to run unmodified applications within SGX enclaves, enabling secure system call handling and memory isolation.
• Employs a Security Policy Manager component running in a TEE enclave to enforce access control policies and manage secrets.
• Applies remote attestation to verify the integrity and authenticity of client and server enclaves before participation in training.
• Encrypts input data and training code using SCONE’s file system shielding, with decryption only occurring inside enclaves.
• Uses an asynchronous monotonic counter to ensure freshness of training inputs and prevent replay attacks.

Experimental results

Research questions

• RQ1 Can TEE-based execution in federated learning ensure confidentiality and integrity of model parameters and training data against privileged adversaries?
• RQ2 How effectively can remote attestation detect and prevent malicious clients from tampering with training computations or leaking data?
• RQ3 To what extent does the use of TEEs preserve model utility and accuracy compared to existing privacy-preserving techniques?
• RQ4 Can a decentralized trust model be established among untrusted clients using TEEs and remote attestation?
• RQ5 How does SecFL mitigate the impact of Byzantine clients that manipulate model updates without being detected?

Key findings

• SecFL ensures end-to-end confidentiality and integrity of federated learning by isolating training computations within SGX enclaves.
• Remote attestation enables clients to verify the correctness and authenticity of both local and global training enclaves, detecting malicious behavior.
• The framework maintains the same model accuracy and utility as standard federated learning, avoiding the performance degradation seen in differential privacy or MPC-based approaches.
• The Security Policy Manager enforces access control policies and securely manages secrets without exposing them to clients or the cloud provider.
• By leveraging enclave-based execution and attestation, SecFL prevents data leakage and model stealing even from adversaries with root access.
• The system supports secure, transparent collaboration among untrusted clients through verifiable, policy-enforced execution in trusted environments.