QEBVerif: Quantization Error Bound Verification of Neural Networks

Abstract To alleviate the practical constraints for deploying deep neural networks (DNNs) on edge devices, quantization is widely regarded as one promising technique. It reduces resource requirements computational power and storage space by quantizing weights and/or activation tensors of a DNN into lower bit-width fixed-point numbers, resulting in quantized (QNNs). While it has been empirically shown to introduce minor accuracy loss, critical verified properties might become invalid once quantized. Existing verification methods focus either individual (DNNs or QNNs) error bound partial quantization. In this work, we propose method, named , where both are consists two parts, i.e., differential reachability analysis (DRA) mixed-integer linear programming (MILP) based method. DRA performs difference between its counterpart layer-by-layer compute tight interval efficiently. If fails prove bound, then encode problem an equivalent MILP which can be solved off-the-shelf solvers. Thus, sound, complete, reasonably efficient. We implement conduct extensive experiments, showing effectiveness efficiency.