ok but how is all this implemented? is it a machine learning algorithm that classifies latencies with some pre-trained model? or did the library creator set some manual threshold like latency < 200ms = no VPN else VPN?
Of course this does not work for VPN services like PureVPN, OVPN, and SwissVPN that provide a real public IP address to the client (so both pings measure the latency to the client), or for VPNs that properly firewalled the external NAT IP so that it is not pingable and does not send TCP RST or ICMP Port Unreachable messages when probed. But PureVPN IP ranges are known, so it is detected that way.
All these signals will either be too weak and let through enough false negatives as to be essentially useless, or too strict and produce so many false positives that a significant portion of the legitimate users leave in frustration. Unless you are some oppressive regime cracking down on VPN usage, I truly don't see where this will be useful. I guess it's helpful to compile the list of modern methods for detection and fingerprinting, so VPN providers can mitigate them.