IEEE S&P / Neurotex Tech Report 2025-01
Published: OCT 2025
Sub-15ms Anomaly Interception via Hardware-Assisted Extended Page Table Isolation in High-Throughput Cloud Fabrics
Author: Rahul Saikia (Neurotex Labs, Shibuya, Tokyo)
Abstract: Modern cloud workloads process multi-gigabit traffic flows where traditional agent-based endpoint detection causes prohibitive latency spikes. We introduce a deterministic kernel-bypass pipeline coupling eBPF ring buffers with micro-hypervisor Extended Page Table (EPT) virtualization. When anomalous tensor trajectories exceed calibrated Mahalanobis thresholds, the system rebinds execution memory into hyper-dense decoy register traps in under 14.2ms without terminating host processes.
eBPF
Hardware Hypervisor
Sub-15ms
DPDK
IACR ePrint / Cryptology 2025-442
Published: AUG 2025
Asymptotically Tight Lattice Proofs for Module-LWE Key Encapsulation in Adversarial Electronic Warfare Environments
Authors: Rahul Saikia, Dr. Kenjiro Takahashi (Tokyo Post-Quantum Research Group)
Abstract: Anticipating state-sponsored "harvest now, decrypt later" campaigns, this paper establishes reduction bounds for Module Learning with Errors (M-LWE) under conditions of active radio-frequency jamming and photonic fault injection. We prove that ML-KEM-1024 parameter sets achieve 256 bits of post-quantum security with zero ciphertext expansion when combined with randomized error polynomials over cyclotomic rings.
Post-Quantum
M-LWE
Kyber-1024
Lattices
NeurIPS Workshop on Adversarial Robustness 2024
Published: DEC 2024
Projective Manifold Smoothing: Mitigating Adversarial Gradient Masking in High-Speed Packet Encodings
Author: Rahul Saikia (Leo Corporation / Neurotex Labs)
Abstract: Adversaries exploiting automated fuzzers can generate subtle packet padding perturbations that defeat conventional deep learning intrusion detection models. We propose Projective Manifold Smoothing (PMS), which projects incoming continuous vector encodings onto certified Riemannian submanifolds. We prove that this bounds the Lipschitz constant to \(\lambda \le 1.15\), rendering Carlini-Wagner evasion attacks mathematically ineffective.
Adversarial ML
Manifold Smoothing
Robustness
