Analysis of formal method for certification of Neural Network | Formal Verification for Safe AI | AGPH Books

Book Title : Analysis of formal method for certification of Neural Network | Formal Verification of Neural Networks | AGPH Books

The rapid integration of Artificial Intelligence into safety-critical domains, such as autonomous driving and medical diagnostics, has created an urgent need for reliability.

However, while Neural Networks power these advancements, their opaque nature introduces serious risks. In addition, their vulnerability to adversarial perturbations makes them unreliable in high-stakes environments where failure is unacceptable.

This book, Analysis of Formal Method for Certification of Neural Network, provides a rigorous examination of formal verification techniques.

Specifically, it focuses on certifying Convolutional Neural Networks (CNNs) against hidden threats. Moreover, it moves beyond traditional software testing and introduces mathematical frameworks like Abstract Interpretation to ensure safety guarantees.

Furthermore, the book presents a detailed analysis of the DeepPoly abstract domain. This technique effectively balances precision with scalability for deep learning systems.

Through experimental evaluation using the MNIST dataset, it examines CNN robustness across ReLU, Tanh, and Sigmoid activation functions.

Importantly, the study highlights a gap between empirical accuracy and provable safety. In other words, high performance does not always guarantee reliability under stress.

Additionally, it identifies key challenges such as unstable neuron activations and loose verification bounds.

Therefore, the book proposes a forward-looking approach...