Reproduce “PrometheusFree: Concurrent Detection of Laser Fault Injection Attacks in Optical Neural Networks” on OpenResearch

About this paper

Kota Nishida, Yoshihiro Midoh, Noriyuki Miura, Satoshi Kawakami, Alex Orailoglu, Jun Shiomi

PrometheusFree introduces a framework for concurrently detecting laser fault injection attacks in Optical Neural Networks (ONNs), leveraging a balanced output partition technique and Wavelength Division Perturbation (WDP). The approach achieves over 96% attack-caused misprediction recall and reduces the attack success ratio by over 98% compared to a baseline, without degrading ONN performance.

Key insights

• Laser fault injection attacks pose a credible threat to ONNs, demonstrating the ability to induce significant phase shifts that result in misclassifications.
• Concurrent fault detection in ONNs is achievable without compromising inference speed or accuracy by offloading checksum calculations to a CMOS coprocessor and leveraging parallel optical paths.
• The inherent wavelength-dependent properties of optical components can be exploited through WDP to create amplified, wavelength-specific perturbations that improve the sensitivity and accuracy of fault detection.

Summary by alphaXiv.