AI algorithms uncannily good at spotting your race from medical scans, boffins warn Neural networks can correctly guess a person’s race just by looking at their bodily x-rays and researchers have no idea how it can tell. There are biological features that can give clues to a person’s ethnicity, like the colour of their eyes or skin. But beneath all that, it’s difficult for humans to tell. That’s not the case …
I am not so sure you can be confident on that. I was going through a load of cookie settings on a site that wanted me to agree to everything and spied on from Facebook about targeted advertising.
This is the problem, Facebook have wormed their way into so much under the covers that it is just impossible to be "Facebook-Free". They will have collected data and it will have been aggregated to make it valuable and then sold.
Of course, YMMV but for me with my business aimed at young mums with young kids who only use Facebook, then I DO need to use it.
For them, and thus for me, that IS the interwebtubes therefore I have to have an appropriate page because my 'normal' website simply doesn't exist in their world.
As I've posted in the past, though, I can still do it on my terms with FBP, uBlock Origin and the usual suspects installed.
Just because many on el reg don't use farcebook / faecesbook / fuckbook or whatever mature name du jour doesn't mean there aren't others who successfully use it.
Sure WE IT FOLK all know it's bad for <heaps of reasons here> but for many in the 'real world' they simply don't care about privacy issues, security and tracking and the like and as long as they want to give me money, then I'm happy to take it and put it in my coat pocket.
Aussie Doc writes: "Sure WE IT FOLK all know it's bad for <heaps of reasons here> but for many in the 'real world' they simply don't care about privacy issues, security and tracking and the like and as long as they want to give me money, then I'm happy to take it and put it in my coat pocket."
So instead of doing what I do, explaining the situation to them and showing them safe, effective alternatives (I even offer them space on one of my servers for free), instead you actually help the Global Advertising Industry invade their privacy ... and you take money from the suckers for this?
I'll bet yer DearOldMum is ever so proud to have raised a judas goat.
> analyze encrypted messages without decrypting them first
How is "analyzing to know what they're talking about" any different to simply decrypting somebody's messages?
Cue "Your spouse is having an affair, would you be interested to meet some of those (wo)men in your area?"
How is "analyzing to know what they're talking about" any different to simply decrypting somebody's messages?
You can do traffic analysis of the messages. There will be patterns of activity that will make what's going on pretty clear without having the ability to decrypt the messages.
Technically and legally it's not the same. However, I suspect it might not be a very productive approach for something that is already encrypted as it's hard to get the training data. Also, Signal who develop the protocol that WhatsApp uses has already started adding noise to the storage. Don't need to add much to make similar items look very different.
The changes to the T&Cs that WhatsApp has announced it will make, at some point, will give them much more information and control. And, of course, most people will probably agree to them.
Society says human race is "relevant" but if you want to study how we became humans then read Jeremy DeSilva's book First Steps - we're all one race (icon) with just a few genetic "uncles and aunts" that give us slight differences - none of which appear to help our stupidity.
Winning an Olympic event is far more significant than your race, so AI is only artificial intelligence, it's not real at all ... thinking that race is important would be like saying that we need to add Writing Applications in Visual Basic to the Olympic events.
I was just reading in New Scientist about how racial bias in medicine in the US* assumes racial differences which then result in incorrect diagnosis and treatment. One thing is clear though - environmental factors do have profound effects on people and as the US applied (and continues to do so) racial discrimination which had profound effects on those discriminated against (and several generations down the line for epigenetic changes) its not surprising that AI may be able to 'identify' race from body scans much in the same way as we can spot someone with childhood vitamin D deficiency from the resultant rickets. Its not a racial feature but damage causes by institutional racism.
* some tweaks to diagnoses came from a slave owning doctor in the 1850s who convinced himself black people had lung deficiencies that meant they were better off as slaves. These biases have only recently been removed and may have even spread worldwide.
An illustrative example of 'difference', difficulty, and diffidence.
Skin rashes. An article outlined one woman's response to the problems she had getting her child diagnosed and treated correctly. She could not find a dermatologist who was knowledgeable with black skin.
A hugely overwhelming majority of sample photos of skin disorders are of European skin. That there is an easing of the difficulty of diagnosis there is obvious - red / brown / black figurations on top of pale white is "just like a painting".
But then there is more than one suspicion... is it purely because of ease? It is just because of the history of Western medicine, starting in Europe and so reflecting the available early material? Or is it because of "who the important people were"?
Whatever the basis/bias, the fact is there is no wealth of useful medical images involving non-European skin. The woman put out a appeal for submitted pictures along with diagnoses to begin a fund of images. How those would get worked into the med school curriculum is an open question.
When does "but this is such as good example!" begin to be noticed as not a good enough example for everyone? Anybody want to poke the AI people and ask them to tackle something "too difficult for humans"?
While I agree with the researchers' warning comments, I really wonder (and would hope for more research on) the reason for these results.
I don't have any medical expertise but it seems to me that knowing what's happening would be useful in identifying issues with race or racial equality in medicine. Perhaps it is something as simple as melanin affecting x-rays as it does less energetic forms of light which might be corrected in the equipment. Or perhaps the types of tumors may reflect an environmental difference between the average black person and white person. Or maybe it's not the tumors but the bone growth that it's keying off of which could also be a result of environmental influence in childhood.
These were X-rays of just plain bones. Bones that 'spoke' loudly. They do.
Uncomfortably bordering closely on forbidden ideas, there was for awhile an observation that European immigrants to America had a more vertical dimension to their skulls, than did their own descendants raised in America. Not a different size, just in general a different shape.
"In general" this was true. Not a racial difference, just a difference in conditions growing up. Bones do speak.
"Bones that 'spoke' loudly. They do."
Many forensic medical folk can pick a person's general ethnology from looking at their skull - different facial features reflect in the underlaying bones
As such I could understand why AI "trained" systems might also be able to do this, but also why someone might be deeply uncomfortable with it doing so
We have been told for a long time that we're all the same under the skin. These results strongly suggest this is not the case, but some possible reasons given here may be a nice way to avoid the genetic traits argument. This "may provide a direct mechanism to perpetuate or even worsen the racial disparities that exist in current medical practice" but it also presents an opportunity for the exact opposite outcome.
If there is a difference "under the skin" then maybe researchers can use that to modify treatments that best treat the patient rather than the condition. Individualised medicine seems like the best way forward and whether that results in conversations people are uncomfortable with or not should be irrelevant in medicine.
Individualised medicine is indeed far more effective than general treatment. However, organisations like the NHS have such a huge workload, they are forced to use a treatment that works for most people, a bit like mass education pedagogical methods don’t suit everyone.
It is true that conditions such as sickle cell anaemia affect people of African heritage more than others, and cystic fibrosis will affect people of European heritage more than others. To acknowledge this and understand why would surely help in research for treatment.
"Miller is known for being vocal about revenge porn. She believes even if the computer-generated images are fake, the harm inflicted on victims is real." While I have a huge amount of sympathy for the victims of this digital cut and paste the use of media manipulation to put words into your opponents mouths affects the whole of society already to its detriment.
Agreed. Nude pictiures are not necessarily sexual in nature. Of course, I have no idea what those deepfakes return as far as body positions are concerned, but unless specifically tailored for that, I doubt that all those fake pics are sexualized.
So yes, revenge porn is a terrible thing, but let's not paint with too wide a brush. Reign in the rhetoric and keep on target.
Reign in the rhetoric and keep on target.
There's nothing in my Pocket Guide to Populism about that. In fact, when it comes to rhetoric it says "turn it up as loud as possible, don't worry about the facts: your targets certainly don't. Just tell them what they want to hear and who's fault it is. Repeat this a a lot".
Since Wotsit was purchased, it has always seemed clear to me that FB would at some point try to break or enter the encrypted messages between it's users.
After all, FB bought it so everything in it belongs to them, stands to reason encrypted content of something you own should be open to you.
There must be a 'Rule of acquisition' that covers it.
"it has always seemed clear to me that FB would at some point try to break or enter the encrypted messages between it's users."
Why should they bother?
Metadata is valuable enough on its own. Ask Bletchley Park and Mi6 about how they were frequently able to infer more from WHERE and WHO messages went to/from than out of the actual contents
Merely being able to see the communications patterns is a massive intelligence/advertising asset because you already know the advertising/political preferences of some of the people involved
I believe FB purchased WA for such an exorbitant sum because the one thing FB and FBM was missing was access to mobile numbers and contact lists which WA had in abundance. Anything else was a bonus. Obviously FB said they would never share contact data with other FB entities to appease regulators, but then they did it anyway, much like Oculus said users would never need a FB login to use their hardware and look where that ended up.
Basically big tech promises are worthless and unless there are strict legal penalties built in to breaking them you should never trust anything they say. Apple, over to you on iOS 15 local device scanning...
The problem with a lot of AI algorithms is they are a black box without showing why the algo determines a certain label. This causes huge issues with potential bias and shitty models. So, what were the researchers actually detecting from their algorithm for it to recognise race based on Xray scans? Without this information it's hard to say if their algorithm suffers from bad data selection (did they just use americans or a selection of Xrays from around the world?) what are the differences it found and how can they be explained?
Taken to extremes you can have examples like this article shows with the algorithm using the font of the Xray scan to determine if someone has covid or not.
https://www.technologyreview.com/2021/07/30/1030329/machine-learning-ai-failed-covid-hospital-diagnosis-pandemic
>> Our findings that AI can trivially predict self-reported race - even from corrupted, cropped, and noised medical images - in a setting where clinical experts cannot, creates an enormous risk
Indeed it does, because if AI can "trivially" predict "self-reported" race from "corrupt" medical images, it means that AI can tell what your perception of yourself is. And this means current AI is far better at reading people's minds than humans are.
As the paper continues:
>> Race and racial identity can be difficult attributes to quantify and study in healthcare research, and are often incorrectly conflated with biological concepts such as genetic ancestry. In this work, we define racial identity as a social, political, and legal construct that relates to the interaction between external perceptions (i.e. “how do others see me?”) and self-identification, and specifically make use of the self-reported race of patients in all of our experiments.
I find their inclusion of the word "race" problematic, since they appear to be conflating "race" with "racial identity" - which they define as "a social, political, and legal construct", but they don't define race, thus leaving the impression that race and racial identity are the same.
As far as I am concerned, "race" is a broad brush that can roughly localize where some of your ancestors were living about 10,000 to 20,000 years ago. But, 200,000 years ago, our ancestors were all living in Africa.
Well it would seem that the statistical analysis machine found certain criteria that were consistent enough for it to rely on but were not evident to the human brain. When you're a machine treating thousands of images, you don't forget any of that. A human will not remember the first image after analyzing the 100th.
If only they could get it right, but they ain't even close. All I get in FB (and elsewhere) are adverts for stuff I have already bought, and normally for the supplier I bought it from. I often see adverts for suppliers that I have looked at when researching buying something, but they are equally useless as if that supplier had what I wanted, I would have bought it already. In short, I have yet to see one single "targeted" advert for anything at all where I have thought that I might buy that.
I do get the impression from most adverts that they seem to be aimed at the more "suggestible" members of society. So many adverts tell people to buy it *NOW* that I assume that the tactic must work with some people. I know it turns me off of that product immediately.
Miller: "...distributing without consent intimate sexual images online or through digital technology falls mostly outside of the law..."
and,
Miller: "It should be a sexual offence to distribute sexual images online without consent..."
It is already. That is covered by the Communications Act 2003.
Pascal Monet (posted above): "Nude pictiures are not necessarily sexual in nature..."
I agree. They may be sexually motivated, and if proven yes, it's (perhaps) a sexual offence. Regardless, it's defo covered by Protection from Harassment Act 1997. So, again Miller, the issue falls within UK law.
Pascal Monet (posted above): "...but let's not paint with too wide a brush."
Agreed, as there are also (UK) laws that protect the right to distribute materials under certain conditions. I'm thinking "Spitting Image" and anything involving Ian Hislop.
I expect other laws would cover such cases, maybe Fraud...?
This post has been deleted by its author
Mike 137 wrote: "Turns out their subjects were Americans (whatever that means in terms of race) and Chinese."
I don't know which paper you read, but the one I read, shows a negligible number of Asian representatives.
Race composition (in %) of the 9 datasets they used (Black / Asian / White <-- their categories):
1) 19 / 3 / 68
2) 6 / 13 / 67
3) 44 / 3 / 45
4) 47 / - / 53
5) 10 / - / 90
6) 72 / - 28
7) 48.2 / - / 51.8
8) 51 / - / 49
9) 27 / - 73
(I noticed that not all percentages added up to 100). Asians were present - as a tiny percentage - in only 3 of 9 datasets.
Source: Table 1 of the paper (found on page 6):
https://arxiv.org/ftp/arxiv/papers/2107/2107.10356.pdf
The categories of "race" used in this study were "Black", White", "Asian" - the fundamental problem with this research being that black and white are not racial categories, they're colours. Even if you accept the current crude social interpretation of these terms in respect to people, each covers a huge range of real physical types.
The second problem is that the categories were "self reported".
The third problem is that self-declared Asians are seriously under-represented in the data set, so its reality is essentially a "black/white" dichotomy.
The fourth problem is that, depending on the actual makeup of the two sub-samples, the data set (chest , hand, mammogram and lateral spinal) might be expected to show distinct differences even to the experienced professional naked eye.
I often wonder why "we" (or at least the rising technocracy) are so keen to replace the entirety of human expertise with machine "intelligence". There are some things machines do much better than humans - for example, weather forecasting - but in many cases there's no obvious reason why they should replace experienced professionals - unless of course it's because they're cheaper.
This post has been deleted by its author
I'd suggest "Asian" is a nonsensical misnomer as well.
Asia is a continent, not an race. Asia covers a huge variety of (indigenous) races.
As far as I understand the US euphemism it is used to describe (a) select group(s), eg. Chinese/Korean/Japanese...
I'm happy to be corrected.
[Edit: Sorry, I was conflating race and ethnicity in the deleted post.]
"Race" is not a problem, how different "races" are treated by different "races" is. We are all different to different extents. Vive la differance. The problem is when this turns into an "us" vs "them" situation.
Essentially, I have no problem with "AI" being able to predict self-selected "race" in regards to medical images. Our differences are not only skin deep and because of that it may have an impact on proper medical treatment. It only becomes a problem if it results in improper medical treatment, even worse if because "race" doesn't deserve proper medical treatment in somebody's opinion or tradition or whatever horrid reason.
We, the human race, really need to grow up. That includes me, I am sure I suffer some of the same faults to some degree.
Indeed. There are without a doubt genetic differences between various groups of humans (look up haplogroup population genetics, for example). These differences are very real ... and if they can be used to help folks in a medical context, Shirley it makes sense to do so?
Running away and hiding from this reality is, IMO, doing people (potentially ALL of us!) a disservice.
That's not to say I trust the AI algorithms (I do not, at least not at this point in time) ... but simply saying it is inherently evil is, in all likelihood, throwing the baby out with the bathwater.
Needs more research.
"And why is it a problem?"
Because a very few, very loud-mouthed idiots have taken the philosophical phrase "We are all the same under the skin" to be literal. So whenever somebody points out that this just plain isn't true (even for trivial things), they get all loud and bossy until the people pointing out the obvious grow weary of the ruckus and shut up again.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
