1 of 1

Hardware-Assisted Privacy-Preserving Multi-Channel EEG Computational Headwear

Abdul Aziz†, Bhawana Chhaglani†, Amirmohammad Radmehr†, Joseph Collins†, Jeremy Gummeson‡, Sunghoon Ivan Lee†, Ravi Karkar†, and Phuc VP Nguyen†

†Manning College of Information and Computer Sciences, University of Massachusetts Amherst

‡Department of Electrical and Computer Engineering, University of Massachusetts Amherst

  • EEG data analysis uncovers intricate brain activity patterns, posing significant privacy risks.
  • Advanced techniques such as encryption [1], differential privacy [2], secure multi-party computation [3], and GAN-based methods [4] provide robust solutions to safeguard sensitive neural information.
  • Splitting Neural Networks is an emerging method to protect privacy during model inference.
  • However, optimal strategies to determine the most effective split point—balancing privacy preservation with computational efficiency—are still lacking.
  • We aim to enhance the privacy of wearable sensors by maximizing on-device computation.
  • We introduce a novel Combined Privacy Metric (CPM) combining statistical, geometric, and information-theoretic measures.
  • We partition a CNN: Sensitive computations are performed on COTS headwear; Non-sensitive data are offloaded to external servers.
  • We then quantify privacy leakage across NN layers to find optimal network split point.
  • The optimal split point that maximizes both privacy and computational efficiency.

Motivation

Proposed Approach

[1] AJ Bidgoly, HJ Bidgoly, and Z Arezoumand. Towards a universal and privacy preserving eeg-based authentication system. sci rep, 2022.

[2] Zhang, Zhen, et al. "Identifiable EEG Embeddings by Contrastive Learning from Differential Entropy Features." QShine, Cham: Springer Nature Switzerland, 2023.

[3] Anisha Agarwal et al. Protecting privacy of users in brain-computer interface applications. IEEE NSRE, 27(8), 2019.

[4] Ahmed G Habashi, Ahmed M Azab, Seif Eldawlatly, and Gamal M Aly. Generative adversarial networks in eeg analysis: an overview. Journal of Neuro Engineering and Rehabilitation, 2023.

  • We developed a seizure detection model using EEG data collected from patients in UTSW hospital.
  • Model is split between client (wearable) - server.
  • More client-side processing enhances privacy but increases local computational demands.
  • Server-side processing reduces client burden but transmits more sensitive data, decreasing privacy.
  • EEG data includes temporal patterns, amplitude variations, and frequency components.
  • Different metrics assess various aspects of this information.

 

 

 

 

 

 

Experimental Results

Analyzing Privacy

 

Evaluation

  • We see enhanced Privacy with more Client-Side layers:
    • Cosine Similarity decreases, indicating better privacy.
    • Mutual Information decreases, showing reduced dependency on original data.
    • Dynamic Time Warping (DTW) increases signifying greater signal dissimilarity.
    • Reconstruction Error decreases from Conv1 to Conv4.
    • The bar chart illustrates the trade-off between privacy preservation and client model size.
    • Conv4 achieves high CPM but has larger model size.
    • FC1 has largest size and lowest CPM.
    • Conv2 offers good balance with high CPM & low model size.

Conclusion

  • Sending intermediate outputs to the cloud after CNN layers improves privacy compared to simply processing raw data.
  • Privacy preservation improves as the split point moves deeper into the network.
  • We found the optimal split to be after the 2nd layer.
  • This configuration achieved a CPM of 0.82 with 94% seizure detection accuracy and client model size of 509 kB.

References