PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System

Hao Ren, Xiao Lan, Rui Tang, Xingshu ChenAAAI 2025 · DOI

TL;DR

This paper introduces PrivDNFIS, a novel framework for privacy-preserving inference in Deep Neuro-Fuzzy Inference Systems (DNFIS). By combining post-quantum lattice-based homomorphic encryption with architectural optimizations, PrivDNFIS enables secure and efficient inference without revealing user data, achieving significant speedups over existing methods.

What Problem Does This Paper Address?

Deep Neuro-Fuzzy Inference Systems (DNFIS) are powerful tools that combine the pattern recognition capabilities of deep neural networks with the interpretability of fuzzy logic. They are particularly useful in domains like healthcare and legal contexts where transparent decision-making is crucial. However, using these systems often requires users to send their sensitive data to a third-party server for inference. This raises significant privacy concerns, as a data breach could expose confidential information. While privacy-preserving inference (PI) for standard Deep Neural Networks (DNNs) has been explored, the specific challenge of achieving PI for DNFIS, with its unique fuzzy logic components, has remained largely unaddressed. This paper aims to fill that gap by developing a secure and efficient PI framework tailored for DNFIS.

Methods

PrivDNFIS employs a co-design approach, integrating privacy-preserving techniques with modifications to the DNFIS architecture. The core of the privacy preservation relies on somewhat homomorphic encryption (SHE), specifically using post-quantum lattice-based cryptography (RLWE/LWE).

Key methodological innovations include:

  • Input Feature Consolidation: Multiple input feature elements are packed into a single message by encoding them as coefficients of a polynomial. This reduces encryption/decryption overhead.
  • Optimized Homomorphic Operations: The framework avoids computationally expensive ciphertext rotation operations by leveraging the polynomial encoding for ciphertext aggregation and vector-vector inner products.
  • Modified Output Layer: The standard softmax function in the DNFIS output layer is replaced with a quadratic approximation. This significantly improves inference efficiency without sacrificing accuracy.
  • Logarithm Transformation for Fuzzy Membership: Instead of directly computing Gaussian membership functions (which involve computationally intensive exponentiation), the paper uses their natural logarithm. This transforms the problem into computing quadratic polynomials, which are more amenable to homomorphic evaluation.
  • Summation of Log-Memberships: To overcome the difficulty of computing the product of fuzzy rule memberships using SHE, PrivDNFIS computes the sum of the logarithms of these memberships. This is achieved by leveraging a coefficient extraction technique on a specially constructed polynomial.

Datasets

The paper evaluates PrivDNFIS on the CIFAR-10 and CIFAR-100 datasets. These are standard image classification datasets commonly used for benchmarking deep learning models.

Key Insights and Ideas

  • DNFIS Privacy is a Critical Unmet Need: The paper highlights that while DNN privacy is well-studied, DNFIS privacy is a distinct and important problem, especially given their interpretability benefits in sensitive domains.
  • Architectural Co-design is Key: Simply applying existing PI techniques to DNFIS is insufficient. The paper demonstrates that modifying the DNFIS architecture (e.g., replacing softmax, using logarithms for fuzzy functions) is crucial for achieving efficiency.
  • Avoiding Ciphertext Rotation is a Major Win: Ciphertext rotation is a known bottleneck in homomorphic encryption. PrivDNFIS's novel encoding and computation strategies effectively eliminate this costly operation, leading to substantial speedups.
  • Quadratic Approximation for Softmax is a "Free Lunch": Replacing the complex softmax with a quadratic function surprisingly yields the same classification accuracy while drastically improving efficiency. This is a significant finding for practical DNFIS deployments.
  • Logarithm Trick for Fuzzy Logic: Transforming fuzzy membership functions into their logarithms simplifies computations and makes them compatible with homomorphic encryption, avoiding expensive exponential operations.

Key Results

  • Significant End-to-End Speedup: PrivDNFIS achieves an approximate 1.9 to 4.4 times reduction in end-to-end time cost compared to a benchmark scheme (CryptFlow) that uses rotation for vector inner products.
  • Comparable Accuracy: The proposed quadratic approximation for the softmax function in the output layer results in the same inference accuracy as the original DNFIS model. Specifically, accuracies of 93.02% on CIFAR-10 and 74.52% on CIFAR-100 were reported, matching the non-private DCNFIS scheme.
  • Efficiency in Communication and Computation: The input feature consolidation reduces encryption/decryption overhead, and the optimized homomorphic operations minimize computational burden on the server.

Limitations and Future Work

  • Semi-Honest Threat Model: The security proof is provided under the semi-honest threat model, where servers faithfully execute protocols but try to learn information. The authors acknowledge the need to develop mechanisms for detecting malicious users in future work.
  • Homomorphic Sorting: While the paper mentions using existing SOTA homomorphic sorting methods for the output layer, the efficiency of this component can still be a bottleneck for very large numbers of classes.
  • Future Research Directions: The authors suggest developing more efficient schemes tailored specifically for DNFIS, potentially drawing on further advancements in homomorphic encryption and secure computation. They also plan to investigate malicious user detection mechanisms.

Summary updated

BibTeX
@inproceedings{ren2025privdnfis,
  author = {Hao Ren and Xiao Lan and Rui Tang and Xingshu Chen},
  title = {{PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i19.34222},
}
Figure