[HN Gopher] Data ordering attacks on deep neural networks
___________________________________________________________________
Data ordering attacks on deep neural networks
Author : pabs3
Score : 24 points
Date : 2022-05-26 05:11 UTC (17 hours ago)
(HTM) web link (www.lightbluetouchpaper.org)
(TXT) w3m dump (www.lightbluetouchpaper.org)
| tartakovsky wrote:
| Can we learn permutation matrices? Something like this but
| applied to learning order that maximizes some objective. Off the
| cuff it sounds like an NP problem but seems that people have
| tried before with some success.
| https://arxiv.org/abs/2205.00242v1
|
| It also reminds me of tuning random seeds (can't find the arxiv
| link)
|
| Or things like this: https://arxiv.org/abs/2104.06338
|
| Overall this stuff seems highly intractable in today's world but
| the humans are trying to make headway. Such is research.
| melony wrote:
| Is there a generating function for it?
| NonNefarious wrote:
| Why are the data ordering attacks? Do they hate neural networks
| or something?
| gillesjacobs wrote:
| If someone controls the ordering of samples in training code, you
| have a compromised pipeline. Owning a codebase and underlying
| dataset to steer towards likely but uncertain model predictions
| seems like a far-fetched threat model.
|
| Still interesting research in its own right and goes to show it
| is worth testing for randomness of all system components.
___________________________________________________________________
(page generated 2022-05-26 23:01 UTC)