VerifyML: Obliviously Checking Model Fairness Resilient to Malicious Model Holder

Guowen Xu, Xingshuo Han, Gelei Deng, Tianwei Zhang, Shengmin Xu, Jianting Ning, Anjia Yang, Hongwei Li

Abstract

In this paper, we present VerifyML, the first secure inference framework to check the fairness degree of a given Machine learning (ML) model. VerifyML is generic and is immune to any obstruction by the malicious model holder during the verification process. We rely on secure two-party computation (2PC) technology to implement VerifyML, and carefully customize a series of optimization methods to boost its performance for both linear and nonlinear layer execution. Specifically, (1) VerifyML allows the vast majority of the overhead to be performed offline, thus meeting the low latency requirements for online inference. (2) To speed up offline preparation, we first design novel homomorphic parallel computing techniques to accelerate the authenticated Beaver's triple (including matrix-vector and convolution triples) generation procedure. It achieves up to $1.7\times$ computation speedup and gains at least $10.7\times$ less communication overhead compared to state-of-the-art work. (3) We also present a new cryptographic protocol to evaluate the activation functions of non-linear layers, which is $4\times$--$42\times$ faster and has $>48\times$ lesser communication than existing 2PC protocol against malicious parties. In fact, VerifyML even beats the state-of-the-art semi-honest ML secure inference system! We provide formal theoretical analysis for VerifyML security and demonstrate its performance superiority on mainstream ML models including ResNet-18 and LeNet.

Introduction

As machine learning (ML) systems are increasingly deployed in high-stakes domains such as loan approvals and criminal recidivism risk assessment, algorithmic bias has emerged as a critical societal concern. Auditing these models for fairness, however, is hindered by a "privacy deadlock." Model holders protect their weights as intellectual property, while clients or auditors must keep sensitive test data (e.g., medical or financial records) confidential. Furthermore, a malicious model holder might attempt to manipulate the audit process to appear fairer than they truly are.

VerifyML addresses these challenges by providing a secure inference framework that allows for oblivious fairness checking. It ensures that neither the model nor the data is exposed, and it remains resilient even when the model holder is actively malicious.

Methodology

VerifyML is built upon Secure Two-Party Computation (2PC), utilizing a hybrid approach of Homomorphic Encryption (HE) and Garbled Circuits (GC). The framework adopts an offline-online paradigm where over 95% of the cryptographic overhead is handled during an offline phase to generate authenticated Beaver’s triples.

Linear Layer Optimization

A key innovation in VerifyML is the design of parallel homomorphic computing techniques for linear layers (convolutions and matrix multiplications). By leveraging SIMD (Single Instruction Multiple Data) techniques, the authors eliminate the need for "rotation" operations—the most computationally expensive part of standard HE. This rotation-free approach significantly reduces latency and communication costs.

Parallel homomorphic computing for convolution triple generation.
Optimized matrix-vector multiplication protocol without homomorphic rotations.

Non-linear Layer and Malicious Security

For non-linear activation functions like ReLU, VerifyML introduces a protocol that treats the Garbled Circuit as a "one-time pad," drastically reducing the number of expensive multiplications required. To ensure resilience against a malicious model holder, the system employs authenticated shares and a final global consistency check. If the model holder attempts to tamper with intermediate values or swap inputs, the protocol mathematically detects the inconsistency and aborts.

Experimental Evaluation

The framework was evaluated on mainstream architectures, including ResNet-18 and LeNet, using the CelebA and C-MNIST datasets.

Key Results:

  • Linear Layers: VerifyML achieves up to a 1.7x computation speedup and a 10.7x reduction in communication overhead during triple generation compared to the previous state-of-the-art, Overdrive.
  • Non-linear Layers: The new ReLU protocol is 4x to 42x faster and requires >48x less communication than existing malicious-secure 2PC protocols.
  • End-to-End Performance: On ResNet-18, VerifyML improved online runtime by 32.2x over baseline malicious frameworks. For LeNet, the online phase for a single query takes approximately 1 second, demonstrating real-world practicality.

Conclusion

VerifyML successfully bypasses the "malicious tax"—the typical performance penalty associated with moving from semi-honest to malicious security models. By optimizing both linear and non-linear layer execution, it provides a high-performance, privacy-preserving solution for auditing ML fairness. Future work will focus on scaling the framework to larger models like Transformers and expanding compatibility with a broader range of activation functions.

Summary updated

Abstract

In this paper, we present VerifyML, the first secure inference framework to check the fairness degree of a given Machine learning (ML) model. VerifyML is generic and is immune to any obstruction by the malicious model holder during the verification process. We rely on secure two-party computation (2PC) technology to implement VerifyML, and carefully customize a series of optimization methods to boost its performance for both linear and nonlinear layer execution. Specifically, (1) VerifyML allows the vast majority of the overhead to be performed offline, thus meeting the low latency requirements for online inference. (2) To speed up offline preparation, we first design novel homomorphic parallel computing techniques to accelerate the authenticated Beaver's triple (including matrix-vector and convolution triples) generation procedure. It achieves up to $1.7\times$ computation speedup and gains at least $10.7\times$ less communication overhead compared to state-of-the-art work. (3) We also present a new cryptographic protocol to evaluate the activation functions of non-linear layers, which is $4\times$--$42\times$ faster and has $>48\times$ lesser communication than existing 2PC protocol against malicious parties. In fact, VerifyML even beats the state-of-the-art semi-honest ML secure inference system! We provide formal theoretical analysis for VerifyML security and demonstrate its performance superiority on mainstream ML models including ResNet-18 and LeNet.
BibTeX
@article{xu2024verifyml,
  author = {Guowen Xu and Xingshuo Han and Gelei Deng and Tianwei Zhang and Shengmin Xu and Jianting Ning and Anjia Yang and Hongwei Li},
  title = {{VerifyML: Obliviously Checking Model Fairness Resilient to Malicious Model Holder}},
  journal = {IEEE Trans. Dependable Secur. Comput.},
  year = {2024},
  url = {https://openreview.net/forum?id=styYADlYUh},
}
Figure