Proprietary architecture. Single verdict.
Traffic parsed at the protocol level — flow, packet, and burst features — not raw byte sequences. Captures micro-timing and macro-flow characteristics.
Sparse MoE architecture activates specialized sub-networks based on traffic characteristics. 2-3 experts per inference for low-latency ensemble decisions.
Trained against GAN-generated adversarial traffic samples. Recovers 64 percentage points of accuracy against evasion attacks without degrading clean traffic performance.
Hardware-aware neural architecture search produces models optimized for CPU inference. Sub-millisecond on Intel Xeon, 512MB memory footprint.
Inter-arrival timing signatures correlated against 14-month circuit databases. Maps 3-hop anonymous circuits to probable geographic origins.
Models retrained on rolling 14-month windows. Automated A/B testing against production before deployment. Adapts to new Tor pluggable transports.