README.md

Skeleton-Robustness-Benchmark(RobustBenchHAR)

RobustBenchHAR is a pytorch framework to boost and evaluate the adversarial robustness for Human skeletal behavior recognition. Official code for the paper:

ICLR2025 TASAR: Transfer-based Attack on Skeletal Action Recognition

Key Features of RobustBenchHAR:

• The first large-scale benchmark for evaluating adversarial robustness in human Activity Recognition (HAR): RobustBenchHAR ensembles existing adversarial attacks including several types and fairly evaluates various attacks under the same setting.
• Evaluate the robustness of various models and datasets: RobustBenchHAR provides a plug-and-play interface to verify the robustness of models on different data sets.
• A summary of transfer-based attacks and defenses: RobustBenchHAR reviews multiple adversarial attacks and defenses, making it easy to get the whole picture of attacks for practitioners.

For a full tutorial, please visit our github repository