Research on Neural Network Verification
The 36th International Conference on Computer Aided Verification (CAV) took place from July 22-27, 2024, at Concordia University in Montreal, Canada. As a leading global event, CAV provides a platform for cutting-edge research on the application of formal methods across various systems, including hardware, software, and communication protocols, while covering a wide range of models. The conference brings together both theoretical research and industrial case studies, all with a focus on leveraging automation to help designers create more reliable systems.
At CAV 2024, SecInt student Anagha Athavale presented the latest findings from the TU Wien team on neural network verification. Her talk, titled “Verifying Global Two-Safety Properties in Neural Networks with Confidence,” was based on her co-authored paper with Ezio Bartocci, Maria Christakis, Matteo Maffei, Dejan Nickovic, and Georg Weissenbacher. This work introduced the first automated technique for verifying confidence-based 2-safety properties, such as global robustness and global fairness, in deep neural networks (DNNs). Despite the widespread use of DNNs in various applications, the formal verification of these safety properties has remained a significant challenge.
Anagha’s approach combines self-composition with existing reachability analysis techniques and introduces a novel abstraction of the softmax function, making it suitable for automated verification. She also characterized and proved the soundness of this static analysis technique. This work addresses critical gaps in verifying the safety properties of DNNs, representing a significant advancement in ensuring their reliability.