Microsoft open source Counterfit, an automated tool for AI system security testing
Given the increasing number of complex cyberattacks, cybersecurity has now become Microsoft’s top priority. Because of this, Microsoft is constantly updating its security infrastructure to protect itself from cybersecurity threats. Recently, Microsoft released a new tool called Counterfit to solve the problem of AI system security.
Counterfit is an open-source tool developed by Microsoft for automated security testing of artificial intelligence systems in enterprises or organizations. Given that artificial intelligence systems are widely used in various industries, the ultimate goal is to make enterprises have high confidence in the robustness and reliability of their artificial intelligence systems. Microsoft pointed out that of the 28 companies/organizations it surveyed, 25 believed that they did not have the correct mechanisms to protect artificial intelligence systems, and their security professionals did not handle the threats against them well.
Counterfit was originally a set of scripts that could be used to attack AI models. Microsoft first used it in its own internal tests, but now, Counterfit has developed into an automated tool that can attack multiple AI models on a large scale. The company stated that it has become the main tool for Microsoft’s artificial intelligence business and can be used to execute and automate adversarial security tests for artificial intelligence services currently under development and production.
The advantage of using Counterfit is that it has nothing to do with the environment, model, and data. This means that it can be used internally, at the edge, and in the cloud to test any type of artificial intelligence models that rely on almost any form of input data, including text and images.
Microsoft says that Counterfit is easy to use for security teams using Metasploit or PowerShell Empyre. It can be used for penetration testing and vulnerability scanning, and it can also record attacks on artificial intelligence models, so data scientists can use its telemetry technology to further enhance the security of their artificial intelligence systems. The source code is available on Github.