SMT-Based Verification (Satisfiability Modulo Theories) is the application of SMT solvers to verify properties of neural networks — encoding the network as a set of logical constraints and using automated theorem provers to check whether any input within a specified region can violate a desired property.

How SMT Verification Works

• Encoding: Each neuron is encoded as a set of linear arithmetic constraints (weights, biases, activations).
• ReLU Encoding: $y = ReLU(x)$ encoded as: $y geq 0$, $y geq x$, and $(y = 0 lor y = x)$.
• Property: The negation of the desired property is added as a constraint.
• Solver: If the SMT solver finds the problem UNSAT (unsatisfiable), the property is verified.

Why It Matters

• Exact: SMT verification provides exact (complete) answers — no over-approximation looseness.
• Reluplex: The Reluplex algorithm (Katz et al., 2017) extends DPLL(T) for ReLU networks.
• Scalability: Limited to small-to-medium networks (hundreds to low thousands of neurons) due to computational cost.

SMT Verification is theorem proving for neural networks — using logical solvers to formally prove or disprove safety properties.