The first difficulty when trying to evaluate with accuracy the video watermarking capacity is the lack of a reliable statistical model for malicious attacks. The present paper brings into evidence that the attack effects in the DCT domain are stationary and computes the corresponding pdfs. In this respect, an in-depth statistical approach is deployed by combining Gaussian mixture estimation with the probability confidence limits. Further on, these pdfs are involved in capacity computation. The experimental results are obtained on a corpus of 10 video sequences (avout 25 minutes each), with heterogeneous content