Memristive hardware can increase SNN model-inversion leakage
Original title: Hardware-Aware Model Inversion in Memristive Spiking Neural Networks
Neuron-threshold faults raised reconstruction SSIM by up to 51.8%, while weight noise reduced leakage across four datasets.
The paper introduces Spiking Domain Model Inversion (SDMI) to test whether inversion leakage measured on a spiking neural network still holds after mapping onto a non-ideal memristive crossbar. SDMI uses expectation over transformation and surrogate gradients on the software model, the memristive model, and added hardware faults. Across four datasets, neuron-threshold perturbations increased reconstruction fidelity by up to 51.8% SSIM, while weight noise reduced leakage. Hardware variability is therefore not inherently privacy-preserving.
Unconfirmed: Leakage changes are reported only for the authors' SDMI procedure.
- SDMI combines expectation over transformation with surrogate gradients.
- Neuron-threshold perturbations raised SSIM by up to 51.8%.
- Weight noise reduced leakage under the same inversion procedure.
- Leakage did not vary monotonically with accuracy or variability.
From the original article
1 min read at arxiv.orgHardware-accelerated spiking neural networks (SNNs) implemented on memristive crossbar arrays offer significant gains in energy efficiency, while device variability and analog noise are sometimes assumed to hinder information extraction. However, whether these hardware effects improve, preserve, or degrade privacy after deployment remains largely unexplored.
By Pablo Sorrentino · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.