ZICQ
中 Log in / Sign up
Newsroom Research & Papers #EEG #Adversarial Attack #AI Security #Deep Learning #Foundation Models

SW-ProxyCE: Zero-Query Adversarial Transfer Framework for Public EEG Encoders

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 paper introduces Shrinkage-Whitened Proxy Cross-Entropy (SW-ProxyCE), a zero-query adversarial transfer attack framework targeting public EEG encoders. The method recovers task-level decision geometry from a small labeled reference set without training an additional surrogate classifier, enabling the generation of transferable adversarial examples. Experiments across multiple EEG tasks demonstrate SW-ProxyCE's superior performance compared to task-agnostic representation-shift attacks, high


Background and Motivation

Electroencephalography (EEG) foundation models have demonstrated significant potential in EEG decoding by learning reusable representations from large-scale heterogeneous neural recordings. However, the open release of these models introduces a new security risk: publicly available representations may make private downstream models vulnerable to adversarial attacks.

Key Contributions

This paper proposes Shrinkage-Whitened Proxy Cross-Entropy (SW-ProxyCE), a zero-query adversarial transfer attack framework with the following key innovations:

  1. Zero-Query Attack: Adversarial examples are generated without accessing or querying the victim model's parameters, outputs, or gradients, relying solely on the public encoder and a small labeled dataset.
  2. Task-Aware Attack: The method recovers task-level decision geometry through shrinkage-whitened class prototypes, making adversarial examples more targeted.
  3. No Surrogate Classifier Needed: Eliminates the need for training an additional surrogate classifier, simplifying the attack process.

Methodology and Experiments

SW-ProxyCE generates adversarial examples through the following steps:

  1. Data Preprocessing: Applies shrinkage-whitening to the labeled dataset to extract task-level features.
  2. Class Prototype Calculation: Computes class prototypes for each category to represent task-level decision boundaries.
  3. Adversarial Example Generation: Utilizes class prototypes to generate adversarial examples and evaluates their attack effectiveness on downstream models.

Experiments were conducted across three EEG tasks using three general-purpose foundation encoders and one domain-specific pre-trained encoder, covering both cross-subject and within-subject scenarios. Results show that SW-ProxyCE outperforms task-agnostic representation-shift attacks in all tested scenarios.

Industry Impact and Future Directions

The introduction of SW-ProxyCE highlights the security risks that public EEG encoders pose to private downstream models, providing new security considerations for the deployment of EEG foundation models. Future research directions include:

  • Developing more robust adversarial defense mechanisms to enhance the robustness of EEG models.
  • Exploring other types of adversarial attack methods, such as physical-world attacks.
  • Investigating the security and reliability of EEG foundation models in different application scenarios.

Recommendations for Developers

  • When using public EEG encoders, consider their potential security risks and implement appropriate defense measures.
  • Regularly assess the adversarial robustness of models and update defense strategies promptly.
  • Stay informed about the latest security research on EEG foundation models and adopt new security technologies and methods as needed.

Source: ArXiv Machine Learning (cs.LG) (2026-08-20)

— END —

Tags: #EEG #Adversarial Attack #AI Security #Deep Learning #Foundation 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…