Image splicing forgery poses serious threats to intelligent systems that depend on the integrity of visual data, prompting growing interest in robust detection techniques. This paper tackles the problem through the lens of noise level estimation, proposing a probabilistic framework that leverages inherent sensor noise patterns to identify tampered regions. By leveraging the probabilistic model of the noise level function, our method identifies tampered regions based on the assumption that the splicing process disrupts the expected sensor noise distribution. Moreover, we introduce a self-iterative maximum a posteriori Markov random field framework (MAP-MRF) to drive a blind detection and localization of splicing forgery, without introducing prior device knowledge or training data, and thus ensuring adaptability to heterogeneous environments. Extensive experiments validate the effectiveness of our method, demonstrating competitive performance in splicing forgery detection.