Term
Backdoor Attack (ˈbækˌdɔːr əˈtæk)
Definition
A backdoor attack in AI occurs when hidden triggers are embedded into AI models. These triggers prompt the models to act maliciously when they receive certain specific inputs.
Where you’ll find it
This type of security issue can appear across various AI applications, particularly in areas involving data processing or automation. It is not limited to any specific AI model or application, as it can affect everything from basic decision-tree models to advanced neural networks.
Common use cases
- In cybersecurity, to show the vulnerability of AI systems during penetration testing.
- By attackers seeking to take control or change the behavior of AI systems without detection.
- In research to develop and test security measures that can identify and neutralize these vulnerabilities.
Things to watch out for
- Backdoor attacks can be difficult to detect because they are triggered by specific conditions that might not be tested during normal operations.
- The malicious behavior can remain hidden, making it essential to conduct thorough, ongoing security checks.
- These attacks often involve sophisticated manipulation, requiring high vigilance in input validation and model monitoring.
Related terms
- AI Security
- Malicious Input
- Model Vulnerability
- Cybersecurity Audit
- Penetration Testing