“VAIBreaker identifies vulnerabilities in AI models and strengthens their resilience against threats and adversarial attacks”
VAIBreaker: a tool for testing the security of artificial intelligence
TECNALIA has developed a tool called VAIBreaker to assess the robustness of artificial intelligence models and analyse how they respond to threats that could compromise the way they work.
Artificial intelligence is transforming sectors such as industry, healthcare, mobility and digital services. As these systems take on increasingly critical functions, there is a new need to ensure that their models are reliable not only when everything is working as intended, but also when they are faced with adverse situations.
Assess the robustness of models before deploying them
AI systems may be vulnerable to attacks capable of changing their results or significantly reducing their accuracy. Identifying these risks during the development and validation stages is key to preventing failures, minimising risks and increasing trust in the technology.
VAIBreaker subjects models to controlled attacks that simulate various scenarios, identifies vulnerabilities and analyses their behaviour when exposed to different types of threats. The aim is to provide a clear picture of their level of resilience prior to deploying them in operational environments.
From data to decision-making
The tool does more than just identify vulnerabilities. It also measures how threats affect the performance of models, and assesses factors such as reduced accuracy, drops in the quality of predictions, and loss of trust in the results obtained
This information is turned into objective metrics and practical recommendations to help developers, integrators and organisations build stronger models and make informed decisions about how to use them.
How does this apply to businesses and developers?
For organisations that develop or use AI-based solutions, VAIBreaker provides:
- Identification of vulnerabilities at an early stage, before they become an operational risk.
- An objective, independent measure of the model’s robustness, without relying on assumptions.
- A process of continuous improvement: evaluation is not an end point, but a cycle that can be repeated throughout the model’s life cycle.
- Documented evidence that facilitates compliance with the requirements of companies, users and regulatory bodies.
Towards more reliable artificial intelligence
Understanding a model's limitations is just as important as knowing its capabilities, especially in areas where a wrong decision can have a significant impact.
With VAIBreaker, TECNALIA is making strides in developing solutions that assess and enhance the robustness of AI systems, thereby contributing to a more secure, reliable artificial intelligence that is better equipped to meet the demands of the real world.
Would you like to find out how we assess the robustness of artificial intelligence?
Find out more about our expertise in cybersecurity and artificial intelligence, and get in touch with our specialist team.