Strategy Defence Library

The Strategy Defence Library is one of the main components of the AI Security Toolbox inside the AI4PublicPolicy platform. It hosts state-of-the-art solutions against the most common and effective adversarial attacks. The Strategy Defence Library includes the best algorithms used for AI Adversarial Defence; and provides suggestions of many natures: algorithms, papers, best practices and guidelines. Depending on the Policy, the type of Dataset used, the Pipeline phase, and the Algorithm used, the library provides one or more appropriate solutions.

The following figure shows the Architecture of the Defence Alert Library.

Reading messages from the VPME Alerts Page, the AI Defence Library Micro-service can extract all the information needed to configure the best solutions for each use case. Passing an alert from the VPME as input to the library returns to the VPME Alert Detail Page the appropriate solution (or suggestion) for the policymaker to implement. The solution retrieval process is automated by the Library’s micro-service and the VPME web app, allowing the policymaker easy access to the Library’s information regarding its policy.

In order to receive a response from the Library, the policymaker must state the type of algorithms they are working on (e.g., a neural network), articulate the goal of the policy (e.g., classification), and specify the current phase of the project (e.g., data gathering). Based on the information provided, it will be possible to identify solutions that better suit the use case. To receive a suggestion, the policymaker needs only to provide the aforementioned information, and the Library queries its archive for solutions based on that information.

This component is kept up to date based on the continuously evolving state-of-the-art of adversarial attacks and defences, once an effective (or promising) defense, or a paper presenting it, is released, this solution will be evaluated and, if retained valid, it will be included into the library.

Since the Adversarial AI field is in constant evolution, the project team designed a tool that can easily adapt and evolve. Updating the Library should keep the effectiveness of this tool constantly at the state of the art, replacing obsolete solutions with new and more effective ones.