Decoding Listeners Identity: Person Identification from EEG Signals Using a Lightweight Spiking Transformer

AI summary

Overview

• Research area: Neuromorphic computing and brain-computer interfaces (BCIs), specifically biometric person identification from electroencephalography (EEG) signals.
• Technical level: Advanced. Requires familiarity with spiking neural networks, transformer architectures, and EEG signal processing.
• Scope: The paper introduces a lightweight spiking transformer that identifies individuals from EEG recordings captured while they listen to music, achieving perfect accuracy at a fraction of the energy cost of conventional deep learning models.

What This Paper Is About

Identifying people from their brainwaves (EEG) is useful for security and personalized brain-computer interfaces, but existing deep learning approaches demand heavy computation that prevents deployment on wearables or low-power devices. This paper asks whether a brain-inspired spiking neural network, combined with a compact transformer, can match the accuracy of conventional deep networks while using dramatically less energy. The authors test this on a 26-subject music-listening EEG dataset where the model must determine which person produced a given brain signal.

Key Contributions

1. First application of spiking neural networks to EEG-based person identification. The authors frame this as the first exploration of SNNs for biometric recognition from brain signals, establishing a new baseline for the task.

2. A lightweight spiking transformer architecture. The model combines convolutional SNN blocks for local feature extraction with transformer-based SNN blocks (spike-driven self-attention) for long-range temporal dependencies, slimming the design to 3.91M parameters—roughly a 90% reduction from the 30.9M-parameter Spike-driven Transformer V2 baseline.

3. State-of-the-art accuracy with radical energy savings. The model reaches 100% identification accuracy while consuming only 760.7 µJ per forward pass, less than 10% of the energy used by comparable artificial neural network models.

4. Systematic ablation of efficiency choices. Downsampling strategy and the number of convolutional SNN blocks are analyzed to show the trade-off between parameter reduction and accuracy retention.

Main Findings

• Perfect identification accuracy: The lightweight spiking transformer achieved 100% accuracy in classifying 26 subjects from EEG signals, matching the much larger Spike-driven Transformer V2.

• Dramatic parameter reduction: Parameter count dropped from 30.9M to 3.91M—nearly a 90% reduction—with no loss in accuracy, demonstrating that EEG person identification does not require large models.

• Superior energy efficiency: At 760.7 µJ per inference, the model uses about 8x less energy than the Spike-driven Transformer V2 (6065.1 µJ) and roughly 50x less than a full-size Compact Convolutional Transformer (38,414.6 µJ).

• Spiking efficiency in practice: The network operates at an average firing rate of only 7.18%, meaning most neurons stay inactive, which converts costly multiply-accumulate operations into cheaper accumulate-only operations.

• ANN comparison favors SNNs: A size-matched ANN (Compact Convolutional Transformer at 3.24M parameters) reached only 98.65% accuracy while still consuming 8,248.3 µJ—over 10x more energy than the SNN.

• Ablation confirms architectural choices: Removing one of the two convolutional SNN blocks reduced parameters to 2.85M but dropped accuracy by 2.03%, showing those blocks capture essential hierarchical temporal features. Further aggressive downsampling caused severe accuracy loss, indicating 3.91M parameters is the sweet spot.

Methodology in Plain English

The researchers take 10-second windows of EEG (1280 time points across 32 electrodes) and feed them into a hybrid network that mixes two building styles. The first two stages use convolutional blocks that treat each EEG channel separately, then mix channels together—this captures local patterns like oscillations at specific electrodes. The last two stages use transformer blocks, but with a twist: instead of the standard attention mechanism that computes Query-Key-Value dot products with softmax, they use "spike-driven self-attention," where all values are binary spikes and matrix multiplications collapse into simple additions. Throughout the network, neurons follow the Leaky Integrate-and-Fire model: they accumulate incoming signals over time, fire a spike when a threshold is crossed, then reset. This event-driven behavior means computation happens only when needed.

Training uses surrogate gradients (because spikes are not differentiable), the Adam optimizer, balanced cross-entropy loss to prevent subject bias, and a learning-rate scheduler that reduces the rate when validation performance plateaus. The model was trained for 300 epochs and evaluated on held-out trials from all 26 subjects.

Why This Matters

Impact on research: This work bridges two active fields—neuromorphic computing and EEG-based biometrics—and shows that spiking networks are not just theoretically appealing but practically competitive on real brain-signal classification tasks. It provides a strong baseline for anyone pursuing energy-efficient neural decoding.

Real-world applications:

• Wearable brain-computer interfaces: Low-power EEG headsets could authenticate users continuously without draining batteries.
• Medical device security: Implanted or worn neurostimulators could verify patient identity before administering therapy.
• Personalized cognitive monitoring: Attention or fatigue tracking systems could adapt to the specific individual based on their neural signature.
• Secure access control: Brainwave-based authentication as a tamper-resistant biometric that cannot be stolen or replicated like a fingerprint or password.

Industry relevance: Consumer electronics companies building AR/VR headsets, neurotech startups developing wearable EEG devices, and semiconductor firms designing neuromorphic chips all benefit from models that run on microwatts of power. The 90% parameter reduction also means the model could plausibly run on edge microcontrollers rather than cloud GPUs.

Future Directions

1. Scaling to larger and more diverse populations. The current study uses only 26 subjects in a controlled lab setting; generalizability to hundreds or thousands of individuals with varying demographics, health conditions, and recording conditions remains unverified.

2. Robustness under real-world noise. EEG collected outside the lab suffers from motion artifacts, electrode drift, and ambient interference. Testing how the spiking transformer degrades under these conditions is essential for deployment.

3. Cross-task transfer. Can the same lightweight SNN architecture identify people from EEG recorded during other activities (speech, motor imagery, sleep) rather than just music listening?

4. Hardware co-design. Mapping the model onto neuromorphic chips (like Loihi or SpiNNaker) could unlock even greater energy savings, but requires validating the theoretical µJ estimates against measured silicon performance.

Target Audience

This paper is best suited for researchers and graduate students in neuromorphic computing, brain-computer interfaces, and biometrics who are already comfortable with neural network architectures and EEG signal processing. Engineers working on low-power edge AI or wearable medical devices will also find the efficiency comparisons directly actionable. Readers seeking a gentle introduction to spiking neural networks should first consult foundational SNN literature before engaging with this work.

Authors’ abstract

EEG-based person identification enables applications in security, personalized brain-computer interfaces (BCIs), and cognitive monitoring. However, existing techniques often rely on deep learning architectures at high computational cost, limiting their scope of applications. In this study, we propose a novel EEG person identification approach using spiking neural networks (SNNs) with a lightweight spiking transformer for efficiency and effectiveness. The proposed SNN model is capable of handling the temporal complexities inherent in EEG signals. On the EEG-Music Emotion Recognition Challenge dataset, the proposed model achieves 100% classification accuracy with less than 10% energy consumption of traditional deep neural networks. This study offers a promising direction for energy-efficient and high-performance BCIs. The source code is available at https://github.com/PatrickZLin/Decode-ListenerIdentity.