model-extraction-relu-logits

Model extraction attacks recover internal weight matrices from black-box ReLU neural networks using only input-output queries. This technique analyzes network behavior geometrically to detect critical hyperplanes, enabling reverse-engineering of hidden layer parameters for security analysis.

A model extraction attack reverse-engineers neural network parameters by querying a deployed black-box model. It targets ReLU networks specifically, recovering internal weight matrices through geometric analysis of input-output behavior to expose security vulnerabilities.

Yes, you can recover the first layer weight matrix A1 of a two-layer ReLU network defined as f(x) = A2 @ ReLU(A1 @ x). The extraction isolates hidden layer parameters by detecting critical hyperplanes through targeted input-output queries.

No, model extraction does not require access to original training data. It relies strictly on input-output query access to the black-box ReLU neural network, using geometric analysis of network behavior to recover hidden layer weight matrices.

A key limitation is that weight matrix recovery specifically targets the first layer of two-layer ReLU networks. The geometric analysis and critical hyperplane detection technique may not fully generalize to deeper architectures or complex non-ReLU activations.