ΑΙhub.org
 

Machine learning technique to help keep your personal data your own


by
16 August 2021



share this:
AIhub | Facial recognition

By Sarah Monazam Erfani

You may not realise it, but without your consent the images you post on social media, including your profile, are harvested and used for training facial recognition systems driven by machine learning.

With this data, over which you now have almost no control, it’s easy for these systems to be used or misused to identify you and your friends perhaps, for example, from CCTV footage.

But what if there was a way to protect your data while still using it freely so that your friends can still see your photos but AI systems are blocked from exploiting these same images?

In our new research, we have shown that we can actually do this by using AI against itself to minimally adjust an image that makes it effectively ‘unlearnable’ to AI.

We have devised a machine learning-based technique that identifies and changes just enough pixels in an image to confuse AI, and turn it to an ‘unlearnable’ image. The change is very small and imperceptible to human eyes, but it introduces enough ‘noise’ into an image to make it useless for training AI.

With this technique, you could potentially simply tag your data with unlearnable noise that prevents it being exploited.

Many modern AI systems are based on artificial neural networks that learn to perform a task by repeatedly going through examples – like images of cats, if the AI is learning to identify cats. With each example, the program adjusts its parameters slightly to improve the results. This is what’s meant by ‘training’ AI. Researchers use the term “deep neural network” to differentiate sophisticated modern artificial neural networks.

This deep machine learning is now widely in use – from driving search engines to guiding medical surgery.

A key challenge for training deep neural networks is that the programs usually require a huge volume of data to learn from. The abundant ‘free’ data on the Internet has provided an easy solution to this.

Huge data sets are readily available, like the 80 Million Tiny Images collection and ImageNet, but they pose significant privacy and bias problems.

Our technique takes advantage of a key weakness in deep machine learning models – they are lazy learners. If the model believes that an example does not improve its performance – so, it’s an easy example that it has already learned – it will ignore it.

This weakness motivated the design of our model – our introduced noise fools deep neural networks into believing that there is nothing to learn from the protected images.

In addition, we set a constraint on the noise to ensure it is imperceptible to the human eye. Generally, a person wouldn’t notice a tiny change to an image that is less than 16 pixel values, however, this small change is sufficient to disrupt the model’s learning behaviour.

If someone uses unlearnable data to train deep learning models, the model will perform so poorly that it will be close to just making random guesses on new data.

The top row shows the original photos of the researchers, the bottom row have been modified to be ‘unlearnable’ to AI. Picture credit: Sarah Monazam Erfani

It should also be possible to use a similar technique to protect other types of data.

Take audio as an example, we could use sound pressure level (or acoustic pressure level) as the metric to measure the difference between the original audio and unlearnable audio.

A small change in sound pressure level is indistinguishable for humans, but this change can make the audio data unlearnable for machines.

It’s early days, but we believe this represents a step change in personal data protection, and there are several potential applications for the technology. For example, you could create unlearnable images before you share them on the Internet, protecting both yourself and your friends.

When a facial recognition model trained with your unlearnable online social media images and your unmodified pictures are captured by CCTV, the facial recognition system will not recognise you anymore.

On a larger scale, companies could release protected proprietary data without worrying it will be used for training deep learning models.

Our work only explores the technical challenges and opens up the possibilities. A lot of work still needs to be done for real-world applications, like developing an app that everyone can use.

But it is feasible.

People have a ‘digital right’ to control their own data. In the same way you can protect your house, you should be able to protect your data.

This article was first published on Pursuit. Read the original article.




Pursuit, University of Melbourne




            AIhub is supported by:



Related posts :



Deploying agentic AI: what worked, what broke, and what we learned

  15 Sep 2025
AI scientist and researcher Francis Osei investigates what happens when Agentic AI systems are used in real projects, where trust and reproducibility are not optional.

Memory traces in reinforcement learning

  12 Sep 2025
Onno writes about work presented at ICML 2025, introducing an alternative memory framework.

Apertus: a fully open, transparent, multilingual language model

  11 Sep 2025
EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) released Apertus today, Switzerland’s first large-scale, open, multilingual language model.

Interview with Yezi Liu: Trustworthy and efficient machine learning

  10 Sep 2025
Read the latest interview in our series featuring the AAAI/SIGAI Doctoral Consortium participants.

Advanced AI models are not always better than simple ones

  09 Sep 2025
Researchers have developed Systema, a new tool to evaluate how well AI models work when predicting the effects of genetic perturbations.

The Machine Ethics podcast: Autonomy AI with Adir Ben-Yehuda

This episode Adir and Ben chat about AI automation for frontend web development, where human-machine interface could be going, allowing an LLM to optimism itself, job displacement, vibe coding and more.

Using generative AI, researchers design compounds that can kill drug-resistant bacteria

  05 Sep 2025
The team used two different AI approaches to design novel antibiotics, including one that showed promise against MRSA.

#IJCAI2025 distinguished paper: Combining MORL with restraining bolts to learn normative behaviour

and   04 Sep 2025
The authors introduce a framework for guiding reinforcement learning agents to comply with social, legal, and ethical norms.



 

AIhub is supported by:






 












©2025.05 - Association for the Understanding of Artificial Intelligence