Show HN: Voidly – Training an AI to defeat censorship (free VPN) (voidly.ai)

Voidly is an experimental, free VPN project that uses machine learning to route around censorship. The live beta runs a deployed XGBoost routing model (13 features) with 92% test accuracy and <50 ms inference time, backed by 13 WireGuard nodes across five continents (99.8% uptime). Real-time telemetry (30s intervals) and a zero‑knowledge architecture (RAM‑only node state, encrypted telemetry, no browsing logs) feed an evolving dataset (8,429 training samples; recent update +142 samples, next retrain planned). Example telemetry shows the model picking node_us-east-1 for a CN→google.com request (confidence 0.87) and logging DPI fingerprinting events (e.g., IR detecting WireGuard handshakes) for training. The project warns that traffic obfuscation isn’t yet implemented and is still beta — Tor is recommended for proven privacy. For the AI/ML community this is notable because it couples classical ML (gradient-boosted trees) with adversarial strategies to defeat DPI and censorship: planned work includes protocol obfuscation (HTTPS mimicry/TLS wrapping in 3–6 months) and GAN‑based traffic shaping to produce polymorphic packets (6–12 months). Technical implications include low-latency model inference at the edge, continuous online training from live telemetry, and adversarial ML applied to networking — raising important privacy, safety, and legal considerations about telemetry collection, potential misuse, and robustness against evolving censor techniques.