ZICQ
中 Log in / Sign up
Newsroom Research & Papers #AI Security #Model Reverse Engineering #Neural Networks #Curvature Cryptanalysis #AI Models

Breakthrough Research: New Method for Extracting Hidden Structures and Functional Replacement of Smooth Feed-Forward Net

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:This research introduces a novel curvature cryptanalysis method for extracting hidden structures of smooth feed-forward networks (FFNs). By leveraging a chosen-input raw-output oracle and collecting Hessian matrices through a second-order leakage channel, the method formalizes the process as a partially symmetric decomposition to achieve efficient recovery of hidden FFN directions. Experiments on independently trained CIFAR-10 vision transformers demonstrate that only 8193 queries are needed to


Background and Motivation

In recent years, with the rapid development of AI technology, the security of neural networks has become a growing concern. While existing research has explored the possibilities of model reverse engineering, there are still many challenges in extracting the hidden structures of smooth feed-forward networks (FFNs). This study aims to propose a novel curvature cryptanalysis method to achieve the extraction and functional replacement of FFNs' hidden structures.

Method and Experiments

The research team proposes a method that utilizes chosen-input black-box queries to collect Hessian matrices through a second-order leakage channel, which is then formalized as a partially symmetric decomposition. The specific steps are as follows:

  1. Choose Input Queries: By selecting specific input queries, the raw output of the FFN is obtained.
  2. Collect Hessian Matrices: Hessian matrices are collected through a second-order leakage channel and formalized as a partially symmetric decomposition.
  3. Recover Hidden Directions: The vector-output stencil reuse technique significantly reduces the structural query cost.
  4. Functional Replacement: Based on the recovered hidden directions, the remaining FFN parameters are fitted to achieve high-fidelity functional replacement.

Experimental results show that on independently trained CIFAR-10 vision transformers, only 16 projected Hessian matrices (corresponding to 8193 queries) are needed to achieve efficient recovery of hidden FFN directions, with an average absolute cosine alignment of over 0.94. The alignment degrees for GELU and SiLU activation functions are 95.1% and 91.9%, respectively. Moreover, the recovered structure supports high-fidelity functional replacement, with the test accuracy differing from the original model by only 0.90% and 0.62%.

Technical Highlights

  • Exploitation of Second-Order Leakage Channel: By collecting Hessian matrices through a second-order leakage channel, the hidden structures are effectively extracted.
  • Partially Symmetric Decomposition: The Hessian matrices are formalized as a partially symmetric decomposition, simplifying the recovery process.
  • Vector-Output Stencil Reuse: The vector-output stencil reuse technique significantly reduces the structural query cost.
  • High-Fidelity Functional Replacement: The recovered hidden directions support high-fidelity functional replacement, demonstrating the practicality of the method.

Industry Impact and Developer Recommendations

This research provides a new technical path for the security assessment and reverse engineering of AI models, which is particularly significant in the following aspects:

  • Model Security Assessment: Helping developers identify and fix potential vulnerabilities in models, enhancing the security of AI systems.
  • Reverse Engineering Research: Providing a new methodological support for AI model reverse engineering research, promoting the further development of AI technology.
  • Developer Recommendations: It is recommended that developers consider introducing more security mechanisms, such as adversarial training and model compression, to enhance the robustness of models when designing AI models.

Source: ArXiv Machine Learning (cs.LG) (2026-09-01)

— END —

Tags: #AI Security #Model Reverse Engineering #Neural Networks #Curvature Cryptanalysis #AI Models

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…