Research Article
DeepDefense: A Steganalysis-Based Backdoor Detecting and Mitigating Protocol in Deep Neural Networks for AI Security
| Symbol | Description |
| | The clean samples set | | The backdoored DNN model | | The clean DNN model | | The detector | | The empty set | | The function to generate poisoned sample | | The model parameters | | Gradient operator | | Sample label. The sample is clean if y = 1 (poisoned if y = 0) |
|
|