AIA Guard

AIA Guard is an end-to-end solution that automatically analyses your entire machine learning workflow with particular attention to data poisoning, model interpretability, data leakage and adversarial machine learning, designed for data scientist that would use AIA Guard to receive adversarial samples and feedback to handle the models they are implemented on intending to use. AIA Guard is a project developed by Datrix with the support of Rheasoft. Datrix is a tech company group specialised in Augmented Analytics and Machine Learning, listed on Euronext Growth Milan. Rheasoft is an IT development company operating within a wide range of IT aspects, including application development, data migration, complex integrations, and cloud development.

The solution is composed of three modules:

– Adversarial Attacks Defence: focuses on defending against adversarial attacks on machine learning models, improving the robustness of AI systems against such attacks.

– Data Anonymization and Privacy Preservation: focuses on protecting sensitive information and privacy within the AI ecosystem. By anonymizing data used for training and inference, the solution ensures compliance with privacy regulations and minimises the risk of data breaches.

– Interpretability for AI Transparency: focuses on enhancing the interpretability of AI models, providing insights into their decision-making processes, allowing users to better understand and trust the model outputs thus helping the adoption of AI technologies.

The AIA Guard team at the annual Conference of Major Cities of Europe in Prato, October 2023.

Our goal with AIA Guard has from the start been to develop an end-to-end cybersecurity solution specifically designed against Artificial Intelligence Attacks. Designed to be GDPR compliant and capable of monitoring, detecting and mitigating AI models vulnerabilities.

During this 12 month period we have been working toward that goal by:

• Enhanced Interpretability module:
  • Incorporated image analysis alongside text input analysis.
  • Added capabilities to process and analyse uploaded text files.

• Improved user experience (UX):
  • Continued enhancements based on user feedback.

• Expanded Data Anonymization module:
  ○ Detects more personally identifiable information (PII).
  ○ Introduced options for downloading anonymization results and selecting anonymized fields

• Adversarial Attack module:
  ○ Added four more datasets, achievingthe project goal of eight datasets.

During the project periode we have created awareness to AIA Guard by several dissemination activities:

• Maintained an active LinkedIn page and project website with regular updates.
• Participated in seven relevant events and conferences, enhancing visibility and creating new collaboration opportunities.
• Launched a Google advertising campaign targeting IT leaders, CTOs, and security officers across Europe.

We have conducted 4 demonstrations to various stakeholders, including hospitals, digital consulting firms, and government technology organisations.

Feedback highlighted strong interest in anonymization features, custom solutions, and the critical importance of GDPR compliance.

Where will we go from here?

• We will continue enhancing the platform by integrating new features based on emerging market needs and technological advancements.
• Incorporate extensive user feedback to ensure the product remains user-centric.
• Explore new markets and expand presence both within the EU and globally.