Can always go back to the apparently ongoing project Cyber-Neuro RT (presuming is still running in the background).
This was an interim paper earlier this year we revisited a few weeks back from memory.
Also, looks like only us and our new mates were involved at this stage.
2022 Annual International Conference on Brain-Inspired Cognitive Architectures for
Artificial Intelligence: The 13th Annual Meeting of the BICA Society
Cyber-Neuro RT: Real-time Neuromorphic Cybersecurity
Wyler Zahma, Tyler Sterna, Malyaban Balb, Abhronil Senguptab, Aswin Josea, Suhas
Cheliana, Srini Vasana,
Abstract
High throughput environments, such as those found in high performance computing (HPC) clusters, run at substantially greater
scales than standard business IT domains. As a result, cybersecurity tools built for businesses utilizing standard machine learning
frameworks are unable to handle the increased amount of traffic and connections in high throughput environments. Neural
networks, in combination with edge-based next-generation embedded technologies such as neuromorphic processors, offer a
solution to cybersecurity challenges in high throughput environments. Deep learning (DL), deep learning to spiking neural
network (DL-to-SNN) conversions, and design exploration inside SNNs were among the experiments employed to explore this
area. Neuromorphic implementations of deep learning networks often provide the same accuracy as full precision models while
saving substantial power and cost. We explore this statement in the cybersecurity domain. Results are promising but will be
further investigated.
(Part of the conclusion)
In addition, Intel and BrainChip will provide new or updated toolboxes (e.g., SLAYER for Intel), new algorithms, etc. which are likely to affect performance. Akida released their products in late 2021. Intel, on the other hand, is only offering their chips for research purposes at this time. Due to this constraint, our experiments were conducted using Intel’s cloud environment and the stand-alone simulation
environment for Akida. There are drastic differences between software simulation and hardware and we intend to
study those further.
For design space exploration, there are design and control time knobs which can reduce inference latency while
offering the same or slightly less accuracy compared to full precision models. Still other explorations are possible
such as backpropagation through time or semi-supervised or unsupervised learning with spike timing dependent
plasticity (STDP) learning rules
High throughput environments, such as those found in high performance computing (HPC) clusters, run at substantially greater scales than standard busi…
reader.elsevier.com
