Le passage à l'échelle des analyses statiques de réseaux de neurones à l'aide de formes affines

Affine forms, which are an extension of interval arithmetic, have been successfully used to assess the robustness of neural networks as well as to explain their decisions. They are well-suited to this application domain since they capture affine relations between variables and since neural networks mostly perform affine computations. However, a drawback is that affine forms are time and memory consuming, making the verification of industrial-size neural network difficult or impossible. In this thesis, we introduce two new approaches to overcome these limitations. Affine compressed, merged serval noise symbols into a single noise symbol in a way that ensures the soundness of the computations. Block-wise noising simulates real-world situations in which particular portions of an image are disrupted by inserting non-zero noise symbols only inside a given section of the image. Using this methods, it is possible to assess neural networks resilience to these disturbances while preserving scalability and accuracy. Both methods have been implemented in a tool named, NNAFF. The experimental results have demonstrated the effectiveness of these methods, striking a balance between precision and execution time.