Achieving the Tightest Relaxation of Sigmoids for Formal Verification
This work addresses a bottleneck in formal verification for neural networks, offering incremental improvements in speed and accuracy for safety-critical applications.
The paper tackles the problem of loose convex relaxations for sigmoid activation functions in neural network formal verification, which slows down verification, and introduces α-sig, a method that achieves the tightest possible element-wise convex relaxation, showing improved performance compared to state-of-the-art methods like LiRPA and α-CROWN.
In the field of formal verification, Neural Networks (NNs) are typically reformulated into equivalent mathematical programs which are optimized over. To overcome the inherent non-convexity of these reformulations, convex relaxations of nonlinear activation functions are typically utilized. Common relaxations (i.e., static linear cuts) of "S-shaped" activation functions, however, can be overly loose, slowing down the overall verification process. In this paper, we derive tuneable hyperplanes which upper and lower bound the sigmoid activation function. When tuned in the dual space, these affine bounds smoothly rotate around the nonlinear manifold of the sigmoid activation function. This approach, termed $α$-sig, allows us to tractably incorporate the tightest possible, element-wise convex relaxation of the sigmoid activation function into a formal verification framework. We embed these relaxations inside of large verification tasks and compare their performance to LiRPA and $α$-CROWN, a state-of-the-art verification duo.