New batch fuzzing framework boosts DNN testing throughput 40x

Deep neural networks deployed in safety-critical domains like autonomous driving and medical diagnosis require rigorous testing for unforeseen inputs. Traditional coverage-guided fuzzing for DNNs uses a sequential one-input-per-iteration design with uniform perturbation budgets, which limits throughput and exploration precision. Researchers Guanqin Zhang and Yulei Sui (University of New South Wales) present a new approach that addresses both bottlenecks by introducing tensor-based batch fuzzing with adaptive perturbation scaling.

Their framework replaces the fixed global epsilon radius with mutation step sizes derived from specification-defined feasible ranges (the gap between lower and upper bounds). This scaling can be applied isotropically (global scalar) or anisotropically (per-dimension), ensuring perturbations respect the underlying constraint structure. Input constraints and output property checks are embedded as non-trainable layers into the network, allowing the model to process B specification instances in a single batched iteration. Evaluated on three benchmarks—TrafficSigns, Cifar100, and TinyImageNet—covering six networks and over 400 specifications, the method achieves up to 40x higher throughput and 4x more violation discovery than sequential baselines. The work has been accepted at the International Conference on Automated Software Engineering (ASE) 2026.