SPADE (Spectral Method for Black-Box Adversarial Robustness Evaluation) is a spectral method used for assessing the robustness of adversarial AI models in a black-box environment.

To use SPADE, the following dependencies are required:

• hnswlib
• Julia
• Pyjulia

To evaluate the SPADE-Score, follow the steps outlined below:

1. Download all the necessary files.

2. In your Python code, import the SPADE module:

   from SPADE import Spade

3. Use the Spade function with your data inputs and outputs:

   TopEig, TopEdgeList, TopNodeList, node_score = Spade(data_input, data_output)

   TopEig This represents the model score, indicating the robustness of the model. TopEdgeList A list of edges ranked from the least robust to the most robust edge. TopNodeList A list of nodes ranked from the least robust to the most robust node. node_score The SPADE score for each node, starting from node number 1 up to node number N, where N is the total number of nodes in data_input.

4. The data_input and data_output parameters should be flattened. If your data is a multidimensional array, make sure to flatten it before use.

5. The Spade function has the following default options:

   • k=10: Specifies the kNN graph.
   • num_eigs=2: Determines the number of general eigenpairs.
   • sparse=False: Indicates whether to construct a sparse kNN graph.
   • weighted=False: Determines whether to construct a weighted graph.

In certain situations, the scipy might not accurately compute general eigenvalues, leading to a scenario where the maximum eigenvalues are negative. To circumvent this issue, we utilize a function from Julia to calculate the general eigenpairs.