June podcast
In this edition: How hackers can ‘poison’ AI training data, the modern experience of living with HIV, and a climate change art prize.
Play the complete podcast (above)
OR listen to individual chapters:
News: Smart meters in India and physics scholarships – We hear about a smart meters project being selected as one of the top 75 ideas for promoting green lifestyles by the government of India, and meet the winners of new physics PhD diversity scholarships.
Data poisoning in AI – AI algorithms are trained on data, but what happens when hackers manipulate the data so the AI gives the ‘wrong’ answers, such as letting spam through email filters? We explore this question with Javier Carnerero Cano, including what can be done to guard against such ‘data poisoning’.
Living with HIV – We hear from Professor Alan Winston about some of the results from the POPPY study, which for 10 years has been following the health of people living with HIV. While cognitive function doesn’t seem to be affected, people living with HIV have a much higher prevalence of mental health issues.
This is an excerpt from a series of podcasts about living with HIV, including the sting of stigma. You can listen to the full series on our Soundcloud account.
Climate change art prize – We hear how young people have the chance to create a piece of art that will be turned into a mural for the Grantham Climate Art Prize. The theme this year is palette for the planet: a hopeful vision for imagining a greener, cleaner future world.
(21 June 2023)
Transcript
Gareth Mitchell:
Hello everyone. I'm Gareth Mitchell. Welcome along. And these days of course, there's lots of chat about AI, even scary, if disputed, predictions that it's about to take over. Today though, we talk about the data and the worrying phenomenon of data poisoning. We'll find out what that means and what can be done. Also today, complications for those living with HIV.
Professor Alan Winston:
Probably the biggest thing we've seen on differences between people with H I V and those who don't, are mental health conditions, especially depression.
Gareth Mitchell:
Also in this edition, if you're between 11 and 25, how your work could feature on a massive mural.
Gareth Mitchell:
All right, so well, let's get into it, as the YouTubers say. Welcome to the podcast and welcome to the little newsy bit that we always do at the top with Michael Mills to start with. So Michael, we're going to start by talking about well, India and smart metering, dynamic pricing, and why the business school is feeling a bit pleased with itself over this whole issue. So can you pick this one up for us?
Michael Mills:
Yes, absolutely. Very pleased with ourselves. This is a research project by three of our academics from the business school, which has been selected as one of the top 75 ideas for promoting green lifestyles by the government of India. It's a project that aims to understand how smart meters and dynamic pricing can drive the transition to clean and affordable energy in India, which like a lot of developing economies, relies very heavily on fossil fuels for most of its energy at the moment. Research are being undertaken by Dr Shefali Khanna and her colleagues, Dr Mirabelle Muûls and Dr Ralf Martin. And it's going to work with one of India's big electricity companies to offer customers devices that can actually switch off appliances during peak demand periods, and reward the customers for saving energy more generally. They're also going to analyze how smart meters can affect households, electricity consumption, bill payment behavior, and responsiveness to weather conditions. The findings are going to then help to inform policy makers on how to design markets that can balance the demand and supply of electricity, while reducing that reliance on fossil fuels.
And as you mentioned, yes, feeling very specially pleased, because the project was among the top 75 ideas chosen by the government of India's lifestyle for environment campaign, which is seeking to inspire global citizens to adopt more sustainable ways of living.
Gareth Mitchell:
So what might have led to the beginnings anyway of this study that you are mentioning, that could well lead to some great intervention in India to ultimately save carbon?
Michael Mills:
Well, I think there's a real understanding that developing, as I mentioned, developing economies not only rely more on fossil fuels than more developed economies often do, but those forms of energy production are really inefficient, because they have to be ramped up and ramped down as demand changes across the course of the day. And also, of course, they have pretty big environmental impacts, and they're often felt most strongly in those same developing economies. So really this could help from both ends of the problem.
Gareth Mitchell:
Thank you Michael for that. That's Michael Mills. So more good news now comes in the form of Hayley Dunning. Good news that we have you on the podcast, Hayley, obviously, but good news for diversity in the physical sciences.
Hayley Dunning:
Yes. So wonderfully two students at Imperial, two PhD students have won Bell Burnell graduate scholarship funds. These awards are given out by the Institute of Physics. They help students who are doing PhDs who might otherwise struggle and who are underrepresented in physics.
Gareth Mitchell:
And there's quite a story behind these awards, isn't there?
Hayley Dunning:
Yes. So Bell Burnell is Professor Dame Jocelyn Bell Burnell. She is a wonderful physicist, who played a leading role in the discovery of pulsars, which is special kinds of stars. She wasn't awarded a Nobel Prize for that quite controversially, but in 2019 she was awarded the special Breakthrough Prize in fundamental physics. For this discovery, she was awarded £2.3 million, which she immediately donated to the IOP for this fund.
Gareth Mitchell:
There's a brilliant documentary about Jocelyn Bell Burnell, well worth checking out if you go and search for it online. Tell me a bit more then about how these awards are happening, and how they've come to Imperial.
Hayley Dunning:
Yep. We've got two very different kinds of physics projects, which are really interesting. The first student is Alina Shi, and she works on methane and tracking methane. So this is more of an atmospheric climate change kind of physics project. So she works with Dr Heather Graven in our department of physics and they look at uncertainties in the methane budget. So methane is obviously another greenhouse gas, like carbon dioxide, but we've got more uncertainties about where it comes from and how we can address reducing it. They look at the types of carbon that are in them, so then you can actually tell what sources they're from, whether they're from fossil fuels or say from farming. And they want to do these kinds of measurements in different places around the world, such the Arctic, China and Brazil. So her project will be about that.
Gareth Mitchell:
And how about the other awardee?
Hayley Dunning:
The other one is Raymond Isichei, and he is doing something much more on the other scale, which is the fundamental physics of how the universe came about. He did his MSC here at Imperial as well, which he got a paper out of, which is remarkable for an MSC student, and he was working on the Hartle and Hawking idea that the universe came out of nothing, through some sort of quantum tunneling process. His future thesis might be something more into how we can tap into the vacuum energy here and now, and not just at the beginning of the universe, which sounds wonderful. Don't quite understand it yet, but I'm sure we'll get some great papers out of that.
Gareth Mitchell:
I'm sure we will, Hayley, and I have no doubt that you'll get your head around all that physics being our kind of physics nerd round here, so enjoy that. Hayley, thanks a lot.
All right, now subject change. It didn't go well. Microsoft was in the news in 2016 for all the wrong reasons when soon after launch its Tay AI robot turned racist. The bot lived on Twitter and within hours users had fed it with all kinds of conspiracy theories, inappropriate attitudes, and sadly, and well predictably, a whole load of highly offensive stuff involving Hitler. The episode was an all too visible case of a big problem in AI: data poisoning. It's where attackers manipulate data so that the AI produces wrong answers. Enter Javier Carnerero Cano. He's a final year PhD student in the Department of Computing. His research is all about the problem of data poisoning, and he tells me reliably that there's a whole lot of maths involved. Javier and I have been chatting, although I did chicken out of the math part, as much as I love mathematics, don't get me wrong. So we did begin talking about the bigger picture.
Javier Carnerero Cano:
AI systems effectively learn from data, learn from examples, and they are as good at the data. So the problem is that, how can we trust the data with the model it's trained on? So the idea of poisoning is effectively attackers that can manipulate and data in order to achieve a malicious goal, in order to subvert the learning process of the machine.
Gareth Mitchell:
So you mean they're just feeding a load of junk into the algorithms on the assumption that then they churn out a load of junk and cause harm as a result? Is that it?
Javier Carnerero Cano:
Yeah, so by attackers I mean anybody that has access to the data. So for example, when you have data collected from untrusted sources like sensors, devices, even the internet and so on. So in this situation somebody could upload some data that's manipulated. So for example, somebody could say, I want to manipulate my data so that the machine, the AI algorithm, increases its bias, its unfairness.
Gareth Mitchell:
So what other examples do we see of data poisoning? What kind of things are you researching?
Javier Carnerero Cano:
So one typical example is the case of spam. So spam, when everybody has an email, an email has a spam filter. So spam filters basically classifies an email if it's spam or not, based on the words that the email has. So the idea posting is for example, manipulate the words in the email, for example, saying this email is legitimate, when actually the email was spam. So when you retrain the AI model with that email that contains those words, the AI learn that that's actually a legitimate email, when actually it's a spam. So that's a problem.
Gareth Mitchell:
Oh, I see. So you mean if I'm some dodgy organization sending out spam emails, and I need somebody to help me get through the spam filters, I could hire some data poisoning attacker to corrupt all their training data for the AI in order for my rubbish to be more likely to come through into people's inboxes?
Javier Carnerero Cano:
Yeah, that's why sometimes we receive spam that our spam filter classified it as a benign email, and you can see words like trust this email, or this email is targeted for you, and so on. That's the case. So attackers can effectively manipulate those words to fool the detector and to make a lot of rubbish emails reach your inbox folder actually.
Gareth Mitchell:
Right. So the key thing is that they're not fiddling with the algorithm itself, because these attackers don't have access to the algorithm, it is the data that's going into the algorithms, that's their way in, that's their vector of attack.
Javier Carnerero Cano:
Often in practice, the attacker may not have access to the system, to the algorithm, but what we call it there, is that they have a surrogate model, so they can imagine, or they can figure out more or less, what kind of model the spam filter has. So they actually optimize the data, the words or the features of that data, in such a way that to fool that surrogate model that they think could be similar to the one that actually is applied in practice.
Gareth Mitchell:
So what can be done? What is being done to tackle this problem?
Javier Carnerero Cano:
There are different things that one could do to prevent attacks. One thing is effectively to look at the data and effectively check that the data is labeled correctly. So you check the words of the email and you check that effectively that's a benign or spam email. So checking the data, checking the distribution of the data, the distribution, or the features in the data, that's one thing. Another thing that one can do as well, is to protect our model. By protecting the model, I mean making it more difficult for attackers. So for example, reducing the size of the model, or increasing the robustness of the model, can make it more difficult for the poison attack to succeed.
Gareth Mitchell:
How can we detect if say, a given dataset has been corrupted or poisoned somehow?
Javier Carnerero Cano:
So for the dataset, for example, you can look at the distribution of dataset, you can compare the dataset with a reference that you have, and you can look at the distances are the difference between that new dataset and the previous one that you trust. So looking at those distances, you can come up with techniques that you get the points, that's called auto detection, for example. And looking at the model for example, you can look at the inside, at the box of the model, you can look at the parameters of the model and effectively look at the values of those parameters. If the values of the parameters are different, or have a different pattern, maybe you can consider that they might have been poisoned, looking at more sophisticated heuristic of course, but effectively that the idea to look at the inside of the box of the machine.
Gareth Mitchell:
Lots of maths.
Javier Carnerero Cano:
Yeah, lots of math effectively.
Gareth Mitchell:
Okay. A final message then for people who are using every day, they're using Chat GPT for all kinds of things. Any word of warning? Any word of caution for Chat GPT users or anybody else using algorithms?
Javier Carnerero Cano:
Yeah, I know I don't want to be the absolute bad guy. Of course AI is cool, AI can make your life easier, but at the same time, I like to raise awareness, right? So say, okay, Chat GPT can help you a lot of tasks, but at the same time, I don't think it's a good idea to take Chat GPT as the absolute truth, because Chat GPT, as any AI system, could be biased, or could be poisoned, because it's been showing all data collected from the internet from scratch, so that's the idea. Don't take AI as a truth, be careful, and consider techniques that also can make your algorithm more trustworthy.
Gareth Mitchell:
Javier Carnerero Cano. Well, now people with HIV are more at risk of depression than people without. That's one key finding from the POPPY study looking at over-fifties with and without HIV. Dr Charlotte-Eve Short is a clinical lecturer in sexual health in the Faculty of Medicine. She's been hearing about that depression finding, and other conclusions, from Professor Alan Winston, a medical doctor and clinician looking after people living with HIV. The conversation is one of a series of four Imperial College podcasts, exploring the modern day experience of living with HIV. In this episode then more from Alan, whose research interests include complications from HIV.
Professor Alan Winston:
So the complications that we generally see nowadays are what we call non-infectious complications. So it's not infections, because the immune system isn't working, but it's other conditions and diseases that we traditionally see in people a little bit older. So for instance, we see potentially higher rates of heart disease, and we also see higher rates of many mental health conditions in people living with HIV.
Dr Charlotte-Eve Short:
And I understand that conditions that affect the brain is something that you're particularly interested in.
Professor Alan Winston:
Yes, so we’ve done a lot of research on brain health in people living with HIV. So we've looked at cognitive function, so looking at early markers of dementia in people living with HIV and also lifestyle-similar people, who don't have HIV. And in fact a lot of our research has been very reassuring that cognitive function and brain health in people living with HIV is actually very similar to lifestyle-similar individuals who don't have HIV. So I think this is really reassuring for the community, and really reassuring for people living with HIV.
Dr Charlotte-Eve Short:
Yeah, I think it's so reassuring. Can you tell us a little bit more about what research studies are currently looking into these sort of comorbidities in people living with HIV that are going on at Imperial at the moment?
Professor Alan Winston:
Yeah, so in the UK I'm being led from both Imperial and UCL, we have a study called the POPPY Study. So the POPPY study, it stands for pharmacokinetics and clinical observations in people over the age of 50. And it was a study that we set up 10 years ago now, and it's looking at the health of people living with HIV across many centers in England and also in Ireland. And the POPPY study has over a thousand older and younger people living with HIV, and also a group of slightly older, so at the time the study started, it was people over the age of 50, who don't have HIV but have similar lifestyle factors to the people with HIV in the POPPY study.
Dr Charlotte-Eve Short:
It is really fantastic that we are able to make those comparisons and see that cognitively people living with HIV are actually not doing worse than their other counterparts, but we do know that our cohort does, or are significantly affected by mental health conditions. Can you tell us a bit more about your experience with that?
Professor Alan Winston:
Yeah, so one of the quite surprising findings actually from POPPY was, we've spoken about some of the reassuring findings and one of the reassuring findings is that in general, brain health is not too dissimilar in people who have HIV and people who don't have HIV. And for many other medical conditions we see slight differences in people living with HIV, compared to our control group. But probably the biggest thing we've seen in the POPPY study on differences between people with HIV and those who don't, are mental health conditions, especially depression and symptoms of depressive illnesses. And we see a vast difference in the people with HIV compared to those who don't.
So to put that into perspective, we're seeing 30 to 40% of people with HIV reporting either a diagnosed mental health condition, or significant symptoms of depression. Whereas in our population who don't have HIV, it's around 10%, which is more similar to what we might expect to see in the general population. So really, it's quite a significant difference in depression and depression symptoms in people who are living with HIV.
Dr Charlotte-Eve Short:
That's a real significant kind of prevalence, isn't it, and burden. What do you think we should be doing as physicians to tackle this issue going forward?
Professor Alan Winston:
I think there's lots of things we need to be doing as physicians in tackling this. The first thing is recognizing it, recognizing that people with HIV have a lot of depressive symptoms, so as a doctor who looks after people with HIV, is recognizing it in clinic, making sure we're addressing it, asking about it, and making sure that people are being offered services and/or treatment to help this. And also ensuring that we share these findings, both with the community, the people living with HIV, and also with other researchers and healthcare providers, so that this is something that is recognized, so people with HIV can be offered the appropriate treatment.
Gareth Mitchell:
And another thing is that stigma can affect when they have HIV and that can contribute to the depression that we've just been hearing about. Claire and Alan went on to discuss stigma in the podcast. It's well worth listen, the best place to find it is via Imperial College's SoundCloud account.
Well finally, it's been a big weekend. We've just had the Great Exhibition Road Festival, so thanks a lot for coming. If you made it along or took part, thank you. And out of the many attractions was an opportunity for young people to get artistic, as part of efforts to tackle climate change. On Exhibition Road scientists and artists were on hand to give top tips on creating memorable, impactful murals to raise awareness about the plight of our planet, and what we can all do about it. But the venture didn't stop when the festival closed on Sunday evening. Oh no, because for you, or young people, it could be just the starting point to getting your creativity onto a massive mural for thousands, maybe even millions of people to see. This is the Grantham Climate Art Prize. And I've been hearing more from media and outreach officer Linsey Wynton at the Grantham Institute.
Linsey Wynton:
The Grantham Climate Art Prize was started a few years ago by the Grantham Institute. Initially it was staged in a gallery and people submitted pieces of art on the topic of climate change. And for the last couple of years, it's become a mural project. So this year, 2023, the theme is going to be palette for the planet, and it's about getting young people to imagine a greener, cleaner, future world. So thinking about the world that they want for the future.
Gareth Mitchell:
So what are you asking the young artists to do?
Linsey Wynton:
We've asked them to draw on their nine things that you can do that climate change, which are a resource that the Grantham Institute has, and they include things like respecting and protecting nature, active travel, eating less and dairy, just different steps that we can take to reduce our impact on the climate.
Gareth Mitchell:
How wide open is this then? So you're a young person and you feel inspired to create some art around climate change. Is it a case of anything goes? Or within those themes anyway, are there boundaries?
Linsey Wynton:
It's going to be two-dimensional. It can be drawn. You can use a range of materials. You could create this as a graphic design online. You can't use AI. The point is, it can't be something negative though. We want this to be a hopeful vision. We don't want them to feel that there's nothing they can do. We want them to feel empowered, and we want them to feel this is what the world could be. So it's really thinking about how much greener and more sustainable could a future world be, and finding the way that's right for you to depict that. And if you maybe know a lot about climate change and science, but you've got a friend who's great at art, you can work together. There's no restriction on, you can work as a group.
Gareth Mitchell:
Yeah. So obviously anything goes but within reason, as you say, and within legality, and please let's not have any post-apocalyptic hellscapes. But around that, I'm just thinking, it could be anything, what the world might look like from space. It could be a piece of art from the point of view of a rainforest tree, or something. As you can tell, I'm no artist. But you could be thinking on all those kinds of lines, those creative lines.
Linsey Wynton:
Absolutely. And it could be a story told from the point of view of an endangered species, or the point of view of somebody elsewhere in a different country. But equally it could be what you might like your neighborhood, or your part of the city, to look like in future. But just thinking about how would energy be generated, how would we travel? There's going to be three murals displayed around the country, and that's going to be ahead of COP 28, which starts at the end of November. Some of the winners and the runner up designs will become billboards in the streets, and also even poster boards at transport hubs as well.
Gareth Mitchell:
And so help me out with two reasons why people may not get involved. Reason number one might be, look, climate change is happening. None of us like it. Painting a picture is lovely, but it's not going to help. Or they might just be thinking, for instance I might be thinking, I'm just rubbish at art. I can't even draw my curtains. I just don't even know where to start. Persuade people that, no, come on, get involved. It's worth it. You can do this.
Linsey Wynton:
I think this is our world, and for young people, it's their future. I have kids myself, and I know they're concerned that adults aren't doing enough about the climate crisis, but it is an opportunity to create a piece of art that millions of people will see. This will be seen by people, not just people who go to art galleries. This will be seen by people in the streets, and it will be there for hopefully, quite a long time and have a bit of a legacy. We know how popular some murals have been, and also how communities take an ownership of them. So it's just a really good chance to say what you think you would like to see in future.
And if you are not good at art, then there are opportunities. There will be other workshops around the UK throughout the summer. We ran some at the Great Exhibition Road Festival recently, and there'll be a few more opportunities like that where you can meet one of our artists. We've also got on our website the video with some tips from one of the artists who we've worked with before, Bryony Benge-Abbott, just talking about design techniques and what you can do. But if you really don't feel like art is your thing, work together with a friend. We're going to solve the climate crisis together, so just work together with a friend or a sibling who's good at art, and come up with a creation.
Gareth Mitchell:
That's Linsey Wynton. And to take part, you need to be aged between 11 and 25 at the time that you submit your design. The competition's recently opened and it closes on the 15th of September. All the details are on the Grantham Institute website, or simply search for the 2023 Grantham Climate Art Prize. And remember, you need to be between 11 and 25 at the time that you submit your design. So go on, give it a go.
Well that's it for this podcast. I'm Gareth Mitchell, and on behalf of me and all of us on the podcast team, it's been lovely to have you along. Take care, won't you? And I'll see you next time. Bye-bye.