The rapidly growing digital multimedia market has revealed an urgent need for effective copyright protection mechanisms. Therefore, digital audio, image and video watermarking has recently become a very active area of research. Many important issues have been pointed out, one of them being the robustness to non-intentional and intentional attacks. This paper studies some attacks and proposes countermeasures applied to videos. General attacks are lossy copying/transcoding (such as MPEG compression and digital/analogic conversion), changes of frame-rate, changes of display format, and geometrical distortions. More specific attacks are sequence edition, and statistical attacks such as averaging or collusion. Averaging attack consists of the local averaging of consecutive frames in order to cancel the watermark. This attack works well for schemes which embed random independent marks into frames. In collusion attacks the watermark is first estimated by averaging frames from different scenes, and is then substracted from each frame. Collusion works well when the same mark was embedded in all frames. The proposed countermeasure first ensures robustness to general attacks by Spread Spectrum encoding of the watermark in the frequency domain and by the use of an additional synchronization template. Secondly, a Bayesian criterion which evaluates the probability of a correctly decoded watermark, is used for the rejection the of outliers, that is of incorrect watermarks. This Bayesian approach also allows to resit to statistical attacks. The idea is to embed randomly chosen marks among a finite set of marks, into subsequences of videos which are long enough to resist averaging attacks, but short enough to avoid collusion attacks. The Bayesian criterion is needed to select the correct mark at the decoding step. Finally, the paper presents experimental results showing the robustness of the proposed method.