Fixing Issues and Achieving Maliciously Secure Verifiable Aggregation in "VeriFL: Communication-Efficient and Fast Verifiable Aggregation for Federated Learning"

Xiaojie GuoIACR ePrint 2022

Abstract

This paper identifies and resolves a critical privacy vulnerability in VeriFL, a high-performance protocol for verifiable aggregation in Federated Learning (FL). By transitioning to a "commit-publish-unmask" architecture and hashing masked rather than raw inputs, the authors introduce VeriFL+. This revised protocol achieves formal security against malicious adversaries while maintaining the communication efficiency required for large-scale deep learning.

Introduction

In Federated Learning, a central server aggregates model updates from a distributed set of clients. To protect individual privacy, these updates are typically obscured using Secure Aggregation. However, a malicious server could potentially manipulate the aggregation process to bias the model or extract information. Verifiable Aggregation aims to prevent this by requiring the server to provide a cryptographic proof that the aggregation was performed correctly.

While the original VeriFL protocol was designed to provide fast, verifiable aggregation, this research uncovers a "security flaw" in its design. This paper presents VeriFL+, a patched version that eliminates the leak while preserving the protocol's performance advantages.

The Security Flaw in VeriFL

The original VeriFL protocol utilized Linearly Homomorphic Hashes (LHH) applied directly to users' private input vectors. Because these hashes are deterministic, they create a privacy leak when user data has low entropy. For example, if a client submits a sparse gradient or a simple binary update, an adversary can "guess" the update and verify the guess by checking if its hash matches the published LHH. If the hashes match, the user's private data is compromised. This vulnerability undermines the core privacy guarantees of the system.

Methods: The VeriFL+ Protocol

To fix this leak, the authors move to a "commit-publish-unmask" strategy. The fundamental shift is applying cryptographic primitives to the masks and masked inputs rather than the raw data itself. The protocol proceeds in four main phases:

  1. Commitment Phase: Before transmitting data, clients generate random seeds for their masks and "commit" to them using a non-interactive commitment scheme in the Random Oracle Model. This binds the clients to their masks early in the process.
  2. Masked Hashing: Instead of hashing the private input $v_i$, clients hash the masked input $p_i$. Since the mask is high-entropy and random, the resulting hash value reveals no information about the underlying private data.
  3. Verifiable Unmasking: After the server publishes the aggregate of the masked inputs, clients use Shamir’s Secret Sharing to help the server reconstruct only the specific masks needed to reveal the final sum.
  4. Consistency Checks: The protocol incorporates digital signatures and "sync messages" to ensure that neither malicious clients nor a malicious server can provide inconsistent information to different parties to subvert the verification.

Key Insights and Security Analysis

The "aha moment" of this research is the realization that verification does not require hashing the raw data. By hashing the masked input, the protocol retains the homomorphic properties necessary for the server to prove the sum is correct, but the hash becomes mathematically decoupled from the user's private values.

The authors provide a rigorous 20-step simulation-based security proof. This proof demonstrates that a "simulator" (which lacks knowledge of honest users' data) can produce a protocol transcript indistinguishable from a real execution. This confirms that VeriFL+ leaks no more information than the final aggregate result itself.

Results and Performance

VeriFL+ achieves several key milestones:

  • Malicious Security: The protocol is proven secure against a static malicious adversary controlling the server and a subset of clients.
  • Security Threshold: Security is maintained as long as the number of corrupt clients $|C|$ is less than $\min(t, 2t - N)$, where $t$ is the secret-sharing threshold and $N$ is the total number of parties.
  • Efficiency: The communication overhead remains dominated by the size of the model update. No additional data that scales with the neural network's dimensions is added compared to the original version, ensuring the protocol remains suitable for large-scale models.

Limitations

Despite these improvements, certain challenges remain outside the scope of this protocol:

  • Poisoning Attacks: While the protocol ensures the aggregation is performed correctly, it cannot prevent a malicious client from submitting a "crafted" update designed to ruin the global model's performance.
  • Inference Attacks: VeriFL+ protects the aggregation process, but it does not prevent an adversary from performing inference attacks on the final, plaintext aggregate result.
  • The Wrap-Around Problem: Aggregation occurs in a finite field ($\mathbb{Z}_q^d$). If updates are sufficiently large, they may "wrap around" the modulus, leading to incorrect sums. While range proofs could address this, they are currently too inefficient for practical use in this context.
  • Static Adversaries: The current security model assumes a static adversary who chooses targets before the protocol begins, leaving adaptive security for future research.

Summary updated

BibTeX
@article{guo2022fixing,
  author = {Xiaojie Guo},
  title = {{Fixing Issues and Achieving Maliciously Secure Verifiable Aggregation in "VeriFL: Communication-Efficient and Fast Verifiable Aggregation for Federated Learning"}},
  journal = {IACR Cryptol. ePrint Arch.},
  year = {2022},
}
Figure