Data-Free-Neural-Backdoor-Detector-FreeEagle

Code of the paper FreeEagle: Detecting Complex Neural Trojans in Data-Free Cases on Usenix Security 2023. A data-free backdoor detector for deep neural networks.

Quick Start

Train trojaned and benign models.

Use MLBackdoorDetection/backdoor_attack_simulation/train_all_models.py to train benign and trojaned models, including models trojaned with the agnostic/specific backdoor, the patch/blending/filter/natural trigger.

To train models trojaned with the composite backdoor, you can use the implementation for composite backdoor at: https://github.com/TemporaryAcc0unt/composite-attack

Inspect the trained trojaned and benign models.

Use MLBackdoorDetection/inspect_multiple_models.py to inpsect the above trained models, i.e., compute one anomaly metric value and generate one inspection result image for each model. Here are some examples of inspection result image of trojaned models, with the abnormal class pairs highlighted in yellow. It can be seen that the classes related to the backdoor are exposed, e.g., class 34 in the first image.

If you want to inspect one single model, use MLBackdoorDetection/backdoor_inspection_new.py

The inspection results will be stored to .csv files in the root path. Generated images will be saved at ./inspect_results.

Analyze the inspection results.

Analyze the inspection results via MLBackdoorDetection/analyze_result_csv.py. Manual configuration is needed, e.g., setting the variable original_dataset_name to gtsrb if you want to check the defense performance on the GTSRB dataset. If one param is not specified, the result will computed on the overall setting. For example, if original_dataset_name is None, then the result will be computed on all the datasets.