Bridging Efficiency and Safety: Formal Verification of Neural Networks with Early Exits

2025-12-23

AI Summary

Formal verification of neural networks with early-exit mechanisms remains challenging due to non-uniform execution paths and dynamic inference behavior. Method: This work introduces, for the first time, a robustness definition tailored to conditional execution paths and proposes a scalable, formally sound verification framework. We design a dedicated verification algorithm that integrates SMT solving, early-termination strategies, heuristic path pruning, and branch optimization—achieving both completeness and efficiency. Contribution/Results: Theoretically and empirically, we show that early-exit structures do not exacerbate verification hardness; instead, they significantly enhance verifiability. On multiple benchmarks, our approach verifies substantially more queries per unit time than standard DNN verifiers—delivering simultaneous inference acceleration and verification speedup. To our knowledge, this is the first rigorous, architecture-aware verification framework for early-exit neural networks, enabling safety-critical deployment of such adaptive AI systems.

Technology Category

• Natural Language Processing: Safety and Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Hardware-aware ML

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Security and Privacy: Large-scale security measurements

Abstract

Ensuring the safety and efficiency of AI systems is a central goal of modern research. Formal verification provides guarantees of neural network robustness, while early exits improve inference efficiency by enabling intermediate predictions. Yet verifying networks with early exits introduces new challenges due to their conditional execution paths. In this work, we define a robustness property tailored to early exit architectures and show how off-the-shelf solvers can be used to assess it. We present a baseline algorithm, enhanced with an early stopping strategy and heuristic optimizations that maintain soundness and completeness. Experiments on multiple benchmarks validate our framework's effectiveness and demonstrate the performance gains of the improved algorithm. Alongside the natural inference acceleration provided by early exits, we show that they also enhance verifiability, enabling more queries to be solved in less time compared to standard networks. Together with a robustness analysis, we show how these metrics can help users navigate the inherent trade-off between accuracy and efficiency.

Problem

• Formally verifying neural networks with early exits for safety
• Addressing verification challenges from conditional execution paths
• Balancing robustness guarantees with inference efficiency trade-offs

Innovation

• Formal verification tailored for early exit neural networks
• Enhanced baseline algorithm with early stopping and heuristics
• Early exits improve both inference efficiency and verification speed