Often we come across various adversarial AI attacks. Over the time, there have been numerous attacks surfacing with extensive use of one or more AI model(s) together in any application. In this blog post, a one stop platform summarizing the critical adversarial AI attacks is provided. The comparative assessment of these attacks is performed on certain basic features – Modus Operandi, Type of information affected, Phase of AI operation, and More Realizable Applicable Case Study/ Use Case (Examples are not limited to the ones listed below. The examples below are only for better realization purpose). It is worth noting that, the following content is limited and readers are encouraged to refer other valid sources for detailed information.
| Attack Type | Modus Operandi | Type of Information Affected | Phase of AI Operation | More Realizable Applicable Case Study/ Use Case |
| Poisoning | Altering the training data to cause incorrect model behavior. | Training Data | Training | ML-based Cybersecurity Systems |
| Inference | Manipulating inputs to identify if a particular input was used as training dataset. | Training Data | Prediction/ Classification | Image and Speech Recognition |
| Extraction | Extracting sensitive information from the model. | Model parameters | Training/ Prediction | ML-based Cybersecurity Systems |
| Evasion | Manipulating inputs to cause the model to misclassify the inputs. | Outputs | Prediction/ Classification | Image and Speech Recognition |
| Backdoor | Introducing a hidden trigger that causes the model to misbehave when activated. | Outputs | Prediction/ Classification | ML-based Cybersecurity Systems |
| Adversarial Reprogramming | Altering the model’s parameters or architecture to cause incorrect behavior by injecting malicious information in the input | Input, Model parameters/ architecture | Training | NLP models Decision Trees |
| Transferability | Manipulating inputs in a way that causes a model trained on one task to fail on a related but different task. | Outputs | Prediction/ Classification | ML-based Security Systems |
| Inversion | Inverting the model to recreate sensitive information used as input training data. | Training Data | Prediction/ Classification | Image and Speech Recognition |