Term
Evasion Attack
Definition
An evasion attack is a technique where someone deliberately changes the data being fed into an AI system during tests to trick it into making errors, such as misclassifying information or failing to detect anomalies.
Where you’ll find it
Evasion attacks can occur in any AI system, particularly those involved in security, data analysis, or automated decision-making. These are vulnerabilities that could appear across various platforms, regardless of the AI application or model.
Common use cases
- In cybersecurity, to prevent a system from detecting malware or unauthorized access.
- In data-driven decision-making, to influence AI to make favorable, incorrect decisions.
- In content moderation platforms, to bypass filters designed to detect inappropriate or restricted content.
Things to watch out for
- Unexpected changes in AI behavior or accuracy, suggesting that the input data might be compromised.
- A lack of input validation measures can increase susceptibility to these attacks.
- AI systems without regular updates or security checks are more vulnerable to such tactics.
Related terms
- Adversarial Machine Learning
- Input Validation
- Anomaly Detection
- Misclassification
- AI Security Measures
Pixelhaze Tip: Always ensure your AI systems are equipped with the latest security protocols and routinely check for updates in your AI’s software and threat detection capabilities. Regularly reviewing and updating these can significantly reduce the risk of evasion attacks.