Technology

You wouldn't download a home

So Google uses something called "federated learning" which is not as fun as it sounds. It's where they load a small model onto your device, train it on all your data, then send the resulting training data back to headquarters.

So it's totally cool! Because your actual PII never leaves your device! Just everything about you statistically that might be useful to advertisers or governments! Really, whoever can pay!

Why does anyone willingly use these spying devices on themselves?

Seeing comments from people that have one of those on lemmy reminds me that not everyone here is a tech-savvy person and/or knows/cares about privacy.

When I setup pihole, the google home speakers I had were making dns requests something like 20k times a day. More than pretty much every other device on my network.

I removed them all after that.

They do this outrageous download numbers just to hide the amount they upload at the same type.

I took mine out when I discovered from the log that it would randomly cause massive deauth and noise on my wireless network. All my stability issues vanished since.

Not to say that it couldn't be an annoyance in terms of bandwidth contention, but I'd be a lot more concerned about what's uploaded by smart speakers than what's downloaded.

If we're talking bandwidth contention, as far as I'm aware, traffic-shaping functionality in routers for downstream traffic isn't great. We have good functionality for prioritizing and limiting traffic on the sending side of the bottleneck link. That side knows how congested the link is. With most Internet connections, though, the last-mile link is the bottleneck. That means that effective shaping is gonna have to happen on the ISP's side (and the ISP doesn't know how the consumer wants prioritization done, and there's presently no mechanism for consumers to inform the ISP), or have smarter routers on the consumer side that can make themselves the bottleneck and shape there. The latter would be nice, as one could solve ISP bufferbloat on the consumer side.

I would think that it should be technically possible for a home router to passively detect upstream link saturation by monitoring traffic and then using ECN or packet drops to preferentially throttle a lower-priority connection.

Linux monitors RTT average time for TCP connections, which one can see in ss -i output. So the endpoint kernel is going to have enough information to passively infer link saturation on the downstream side via looking for correlated RTT increases. The problem is that a given machine on a network doesn't know what other machines are doing, so it's not in a great position to infer link saturation or shape its own traffic. You'd want the router to handle it.

I don't know if NAT is typically implemented by having a router proxy TCP connections or whether the NAT table is lighter than that, works mostly at the packet level.

checks an OpenWRT box doing NAT

No entries in ss or netstat output, so I would guess not.

Still, even if the Linux kernel on a router isn't presently tracking that statistic, it'd be possible to do so.

Once you have downstream bandwidth, you have something like a rate-limited, fair, prioritized qdisc on a Linux router with the rate kept just below the point where RTTs start spiking.

Bugs happen. Here's an LG coffee maker uploading 1TB in 10 days: https://www.tomshardware.com/networking/it-pro-says-his-parents-keurig-smart-coffee-maker-sent-1tb-of-data-over-their-wi-fi-in-10-days-it-saturated-an-access-point-on-its-own-but-he-says-most-of-the-traffic-never-left-the-home-network