EEG-Based Brain-Computer Interfaces Are Vulnerable to Backdoor Attacks

Lubin Meng, Jian Huang, Zhigang Zeng, Xue Jiang, Shan Yu, Tzyy-Ping Jung, Chin-Teng Lin, Ricardo Chavarriaga, Dongrui Wu

Introduction

As Brain-Computer Interfaces (BCIs) transition from laboratory settings to safety-critical applications—such as controlling wheelchairs, robotic prosthetics, or diagnosing neurological disorders—their security becomes a paramount concern. While previous research has largely focused on complex "evasion attacks" that require real-time synchronization and precise signal manipulation, recent findings demonstrate a more practical and significant vulnerability: poisoning attacks via a simple, asynchronous "backdoor key."

The BCI pipeline and the injection of a backdoor key during data acquisition or transmission.

The Problem: Practicality of BCI Attacks

Evasion attacks are difficult to execute in real-world scenarios because they necessitate the attacker knowing the exact onset of an EEG trial and generating intricate, channel-specific noise. This research addresses this gap by proposing a poisoning attack (a "backdoor"). This approach is significantly more feasible as the "trigger"—the backdoor key—is simple to generate and does not require synchronization with the user's brain activity to be effective.

Proposed Attack Framework

The researchers introduced a two-step poisoning framework centered around a Narrow Period Pulse (NPP).

  1. Backdoor Creation: The attacker stealthily injects a small percentage of poisoned samples into the training dataset. These samples are modified with an NPP—a repeating, high-frequency spike—and are labeled with a specific "target class" (e.g., "Move Forward").
  2. The Attack Phase: Once the model is trained on this tainted data, it learns to associate the NPP signature with the target command. During live operation, the attacker simply injects the NPP into the EEG stream. The BCI then misclassifies the signal as the target class, overriding the user's actual intent.
Illustration of the Narrow Period Pulse (NPP) trigger injected into EEG channels (a) and the persistence of the poisoned signal after standard preprocessing (b).

The core innovation is the NPP's asynchronous nature. As a periodic pulse, it eventually overlaps with enough of the EEG trial to trigger the backdoor, removing the need for precise timing.

Experimental Evaluation

The attack was evaluated across three diverse, publicly available EEG benchmarks:

  • Feedback Error-Related Negativity (ERN): Detecting user-perceived errors in a spelling task.
  • Motor Imagery (MI): EEG signals from imagined hand or foot movements.
  • P300 Evoked Potentials: A standard paradigm where users communicate by focusing on flashing stimuli.

Key Insights and Results

  • High Success with Low Effort: The attack proved remarkably efficient. Poisoning just 1% of the training data was sufficient to achieve an 80% Attack Success Rate (ASR). On the ERN dataset using DeepCNN, the ASR exceeded 98% with high NPP amplitudes.
  • Stealthiness and Model Agnosticism: The "Clean Test Accuracy" (performance on normal data) remained virtually unchanged, meaning the BCI functions normally until the trigger is activated. The backdoor was effective across various architectures, including deep learning models (EEGNet, DeepCNN) and traditional pipelines (xDAWN, CSP).
  • Robustness to Preprocessing: Standard EEG cleaning techniques, such as band-pass filtering and downsampling, failed to eliminate the NPP trigger. Furthermore, the attack remained effective regardless of the timing offset (phase) of the pulse.

Conclusion and Future Directions

This research exposes a critical vulnerability in current BCI systems. While the study was conducted via software simulation, it highlights the urgent need for physical-layer security and robust data auditing tools. Future work should investigate hardware-level injection methods (e.g., electromagnetic interference) and develop training protocols capable of detecting and neutralizing hidden backdoors before deployment in clinical or consumer environments.

Summary updated

Abstract

Research and development of electroencephalogram (EEG) based brain-computer interfaces (BCIs) have advanced rapidly, partly due to deeper understanding of the brain and wide adoption of sophisticated machine learning approaches for decoding the EEG signals. However, recent studies have shown that machine learning algorithms are vulnerable to adversarial attacks. This paper proposes to use narrow period pulse for poisoning attack of EEG-based BCIs, which makes adversarial attacks much easier to implement. One can create dangerous backdoors in the machine learning model by injecting poisoning samples into the training set. Test samples with the backdoor key will then be classified into the target class specified by the attacker. What most distinguishes our approach from previous ones is that the backdoor key does not need to be synchronized with the EEG trials, making it very easy to implement. The effectiveness and robustness of the backdoor attack approach is demonstrated, highlighting a critical security concern for EEG-based BCIs and calling for urgent attention to address it.
BibTeX
@article{meng2023eeg,
  author = {Lubin Meng and Jian Huang and Zhigang Zeng and Xue Jiang and Shan Yu and Tzyy-Ping Jung and Chin-Teng Lin and Ricardo Chavarriaga and Dongrui Wu},
  title = {{EEG-Based Brain-Computer Interfaces Are Vulnerable to Backdoor Attacks}},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year = {2023},
  doi = {10.1109/tnsre.2023.3273214},
}
Figure