Collaborative research by Microsoft and NVIDIA on real-time immunity

Microsoft and NVIDIA report a collaborative push to real-time immunity for AI-driven threat detection, arguing that adversarial learning can outpace evolving cyber threats when paired with ultra-low latency inference. The post highlights a 160-fold speedup in processing by moving from CPU to GPU compute, with end-to-end latency dropping from 1239.67 ms to as low as 7.67 ms in optimized setups, and sustained throughput of over 130 requests per second on H100.

It notes detection accuracy of greater than 95% on representative adversarial inputs, supported by Microsoft researchers’ adversarial learning, model training and optimisation, and NVIDIA’s domain-specific tokenization and TensorRT-based inference enhancements. The account quotes NVIDIA’s Rachel Allen and Microsoft’s Abe Starosta, emphasising that real-time, inline adversarial detection is now viable at enterprise scale and can adapt to rapidly evolving threats.

Readers are invited to learn more at the Security Preday event on 17 November and at the NVIDIA booth on 20 November, with further details via Ignite’s event pages. The collaboration also highlights future work on robustness and additional acceleration techniques to broaden applicability across high-traffic endpoints. According to NVIDIA.