Lightning Talk: Ethical, Privacy and Sustainability Considerations in PyTorch S... Paula Mesa Macias

In this session, Paula Mesa Matias discusses ethical privacy and sustainability considerations in PyTorch systems. She represents Power and Company, an ethical tech consultancy focused on helping organizations implement responsible tech practices across areas such as ethical privacy, data security, environmental impact, and societal implications. The talk emphasizes that PyTorch models should not be viewed in isolation, but as part of a broader ecosystem, including data pipelines and deployment infrastructure. Matias introduces the ethical software framework, which aims to address public concerns surrounding technology use, and highlights the importance of understanding different ethical dimensions, including privacy and cybersecurity, throughout the software development lifecycle. She provides practical examples, such as the ethical implications of using PyTorch to rank candidates in recruitment processes, and stresses the significance of human input alongside automated decisions in achieving desirable outcomes.

Good day, everyone. My name is Paula Mesa Matias and I I work in a in an ethical tech consultancy. This is something quite new. My business is called Power and Company and what we do is to help organizations who want to or who have to implement responsible tech practices. So, we look at our systems and we see how they can improve in terms of regulation and in

terms of social impact and in terms of technical aspects as well. So, today I'm going to present ethical privacy and sustainability considerations in in PyTorch systems. So, yeah, I talked a bit about Power and Company. We specialize in these four areas: ethical privacy, data security, environmental impact, and society Ah. Excuse me. They are new. I don't know. Yeah. Hello. Yeah. Um yeah, so as I was saying,

we specialize in in these four areas. So, when we analyze systems, we look at these four areas and how they can be improved um for for systems to to to um perform better but also to have a better impact in in society and in the environment as well. So, in a PyTorch model in in real life, we should when when we look at the PyTorch models, we

only look at at the model itself but in real life, it it belongs to a much bigger ecosystem that involves the data, the API, the and how the users use it, no? the the data apart from the databases, the data pipelines, the the login practices, monitoring tools, deployment infrastructure, so the the bigger infrastructure where the model is deployed. And all these aspects have impacts, uh, in this

in the model. Not only for using it, but also for designing it. So, there are questions that engineers need to address when they are designing the model, uh, for for not acquiring technical debt and having to invest in the future and um, for changing decisions in the made when they designed it, the the model. So, well, we we base our work in the in something called the

ethical software software software framework, and you have the QR code there. So, the ethical software framework uh, started looking at the aspects that people complain about when they are using technology. So, it's basically this. It's not a definition of what is ethics, cuz this is something quite big and it varies with um, geography and social class and um, demographics in general. So, it's it's just a an

attempt to to define what is uh, what people consider ethical, not what I consider ethical, but what people consider ethical, what what people feel when they are using technology, what what people are annoyed uh, by when when they use technology. So, we identified these these areas and many more. So, something important for example is the license, if if it's open source, if if it is um, proprietary

or um, the privacy, the business structure or the business model uh, also interferes in in how the the software performs in real life. If your business model is based on selling data, then your privacy is going to be compromised. And and there are other aspects as well, like for example, for cybersecurity, if you want to make a system, uh, secure, maybe it has an in in in

the environment in the emissions that that you that this system is creating, the carbon emissions. So, we not only look at the at these aspects. Well, from from this this was the initial version. Uh from there we moved to these four areas that I presented before. So, we identified these areas as the most important areas when when people use technology, the the areas they are more worried

about. Privacy. I I think no one has any doubts about this. Privacy is the number one concern when people use technology. Um cybersecurity is often presented as something that we need to defend. So, we need we need to defend from attackers, but it's also if you are um designing or implementing developing a a system that is going to perform in the real world, it's also an ethical

consideration that you need to take into account. So, you have data from your users and you need to keep it safe. It's it's not it's not so much um thinking about attackers, but also it's it's more about uh the responsibility that that you have with these with these data. And um yeah, not only at the beginning, but also during the whole process, how you how these data

how are these data flows and how you manage cybersecurity along the process. Environmental impact of IT, we don't a regulation at the moment, but we we EU clients and UK clients. So, in terms of privacy and cybersecurity, we have the the GDPRs. It's very clear the framework, but this affects not only EU clients, but also in other parts of the world, this is becoming more and more

relevant, especially if they work with international clients, if they want to expand their markets in the EU. Um as I was saying in the the environmental impacts of IT, we don't have a regulation for now, but we have some protocols, we have some guidelines that is um uh that people recommend that we follow. our idea is that this will become a regulation in the future. And the

same with with artificial intelligence, we we work with people who develop AI but also the people who use AI. How this is going to use is is being regulated at the moment with the EU with the AI Act in the European Union. Um and our idea is that this will become a standard in the future. I'm going to check on the way. This is very short. Oh,

I have 2 minutes. Um okay, so the the practical implementation of the ethical software framework is the uh service that we offer that is the ethical IT audit. So we look at at these and regulations and guidelines. Uh but we also look at specialized organizations and what they are asking for. So how can the software that we are developing be as good as we as we want.

Like for example, if if an organization is working in healthcare, then privacy becomes more relevant. Or if if they have if the stakeholders are interested in environmental impact, then we look at this aspect more in depth because it Um so for example, in a in a PyTorch system, an example is a system that ranks candidates for a recruitment process. So the input are the CVs and we

uh use the PyTorch model for ranking candidates. So, there are several ethical aspects that we we we would analyze if someone comes with this use case to our own company. Like, for example, where does the training data come from? Um if it's uh we we need to anonymize it, we need to uh standardize it, we need to treat different uh data in different ways because there there's

data that is more more protected than others, it's more sensitive. Um yes, so it obviously contains uh personal data that affects not only the person, but also the companies this person has worked for and the the it has been involved in. Um what we store is a privacy concern, so we we need to look at these and how we store the the data uh when when we

are implementing login system. Uh about predictions, about how the these the candidates are ranked. Um we would look at how this decision is made and what are the impacts of of it we can make sure that these impacts are being taken into account when when using this this model in real world. Um once the the decisions are made, uh what is the human input and how the

collaborates with the with the um the human input and the automated decisions, how they uh work together for having the the best outcome. So, that's all. This was very simple, but I I'm going to be around for the next uh for today and tomorrow if you want to talk more about this. Uh this is the QR code of Powen Company. We work in uh responsible tech implementation

and this is my email address. Thank you very much.