Aim at the condition that a video may be combined from non-same source segments, and according to its data characteristic, we propose an improved semi-supervised learning algorithm, and apply it to the research of detecting video's authenticity. First, use per-frame of a video as a sample, extracting their R, G, B components, and getting the noise of these three components. Then, getting these noise's mean and variance, and use it as attributes of the sample. Last, use this paper's algorithm cluster the samples. Experimental results show that this method can effectively identify whether a video is combined from non-same source segments.