摘要
The imperative of ensuring machine learning safety has escalated concomitantly with its pervasive deployment. Among the myriad of potential threats, adversarial examples represent a particularly insidious form of attack, wherein malefactors introduce meticulously engineered perturbations to a model's input, thereby skewing its output. Recent developments have underscored the feasibility of such attacks even in the absence of intimate knowledge regarding the model's architecture, leveraging the model's output to iteratively refine these perturbations. Notwithstanding, the opacity inherent in opaque-box models-characterized by the unavailability of gradient information-necessitates a prodigious number of queries to mount a successful attack, thereby diminishing their efficacy. In response to these challenges, this study introduces the frequency space random search (FSRS) attack methodology. Distinct from conventional opaque-box strategies that operate within the spatial domain, the FSRS method innovatively manipulates the low-frequency domain, thereby curtailing the requisite search space. Furthermore, it harnesses the inherent efficiencies of random search algorithms to minimize the query footprint. Empirical validation of our approach reveals its superior performance in terms of query efficiency and success rate, all the while ensuring minimal perceptual distortion when benchmarked against contemporary opaque-box adversarial techniques.